STARFlow: Scaling Latent Normalizing Flows for High-resolution Image Synthesis
Extracted source text · arXiv 2506.06276
\documentclass{article}
% if you need to pass options to natbib, use, e.g.:
% \PassOptionsToPackage{numbers, compress}{natbib}
% before loading neurips_2025
% ready for submission
%\usepackage{neurips_2025}
% to compile a preprint version, e.g., for submission to arXiv, add add the
% [preprint] option:
\usepackage[preprint]{neurips_2025}
% to compile a camera-ready version, add the [final] option, e.g.:
% \usepackage[final]{neurips_2025}
% to avoid loading the natbib package, add option nonatbib:
% \usepackage[nonatbib]{neurips_2025}
\usepackage{xcolor}
\usepackage{enumitem}
\usepackage[utf8]{inputenc} % allow utf-8 input
\usepackage[T1]{fontenc} % use 8-bit T1 fonts
%\usepackage{hyperref} % hyperlinks
\usepackage{amsmath}
\usepackage[most]{tcolorbox}
\definecolor{mydarkblue}{RGB}{0, 2, 115}
\definecolor{cutelightblue}{RGB}{173, 216, 230}
\definecolor{cutelightred}{RGB}{255, 182, 193}
\newcommand{\highlight}[2]{\colorbox{#1}{$\displaystyle#2$}}
\usepackage[colorlinks=true,linkcolor=mydarkblue,citecolor=mydarkblue,urlcolor=mydarkblue,pdfborder={0 0 0}]{hyperref}
\usepackage[capitalize]{cleveref}
\crefformat{equation}{Eq.~(#2#1#3)}
\crefname{section}{§}{§§}
\Crefname{section}{§}{§§}
\crefname{remarkbox}{Prop.}{Props.}
\Crefname{remarkbox}{Prop.}{Props.}
\usepackage{url} % simple URL typesetting
\usepackage{booktabs} % professional-quality tables
\usepackage{amsfonts} % blackboard math symbols
\usepackage{nicefrac} % compact symbols for 1/2, etc.
\usepackage{microtype} % microtypography
\usepackage{xcolor} % colors
\usepackage{graphicx}
\usepackage{wrapfig}
\usepackage{multirow}
\usepackage{wasysym}
\usepackage{caption}
\usepackage{subcaption}
\usepackage{amssymb}
\usepackage{wrapfig}
\usepackage{listings}
\usepackage{pifont}
\usepackage{caption} % gives \captionof
\usepackage{booktabs} % already needed for \toprule etc.
\usepackage{amsthm}
\usepackage{placeins} % optional, for \FloatBarrier
\tcbuselibrary{breakable}
% \newtcolorbox{remarkbox}{
% enhanced,
% breakable,
% colback=gray!10,
% colframe=gray!50!black,
% coltitle=black,
% %title={\textbf{Remark}},
% fonttitle=\bfseries,
% left=1mm,
% right=1mm,
% top=1mm,
% bottom=1mm,
% boxrule=0.5pt,
% arc=4pt,
% }
\newcounter{remarkbox}
\renewcommand{\theremarkbox}{\arabic{remarkbox}}
\newtcolorbox{remarkbox}{
enhanced,
breakable,
colback=gray!10,
colframe=gray!50!black,
coltitle=black,
fonttitle=\bfseries,
left=1mm, right=1mm, top=1mm, bottom=1mm,
boxrule=0.5pt,
arc=4pt,
% BEFORE the box body: step the counter, make it ref-able, and print the title
before upper={%
\refstepcounter{remarkbox}%
\noindent\textbf{Proposition~\theremarkbox.}\quad
},
}
\definecolor{deepblue}{rgb}{0,0,0.6}
\definecolor{deepred}{rgb}{0.6,0,0}
\definecolor{deepgreen}{rgb}{0,0.5,0}
\lstdefinestyle{python}{
language=Python,
basicstyle=\ttfamily\small,
commentstyle=\color{deepred},
otherkeywords={}, % Add keywords here
keywordstyle=\color{deepgreen},
emph={}, % Custom highlighting
emphstyle=\color{deepblue}, % Custom highlighting style
% stringstyle=\color{deepred},
% frame=tb, % Any extra options here
showstringspaces=false %
}
\input{commands}
\input{math_commands}
\newcommand{\Model}{\textsc{StarFlow}}
\newcommand{\model}{STARFlow}
\newcommand{\pnf}{p_{\rm{NF}}}
\newcommand{\pam}{p_{\rm{AR}}}
\newcommand{\paf}{p_{\rm{AF}}}
\newcommand{\pencoder}{q_{\rm{enc}}}
\newcommand{\pdecoder}{p_{\rm{dec}}}
\newcommand{\tc}[1]{\textcolor{blue}{[TC: #1]}}
\title{\ding{72}\Model{}: Scaling Latent Normalizing Flows for High-resolution Image Synthesis}
% The \author macro works with any number of authors. There are two commands
% used to separate the names and addresses of multiple authors: \And and \AND.
%
% Using \And between authors leaves it to LaTeX to determine where to break the
% lines. Using \AND forces a line break at that point. So, if LaTeX puts 3 of 4
% authors names on the first line, and the last on the second line, try using
% \AND instead of \And before the third author name.
\author{%
Jiatao Gu, Tianrong Chen, David Berthelot, Huangjie Zheng, Yuyang Wang, Ruixiang Zhang, \\
{\bf Laurent Dinh, Miguel Angel Bautista, Josh Susskind, Shuangfei Zhai} \\
Apple\\
\texttt{\{jgu32, szhai\}@apple.com} \\
% examples of more authors
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \AND
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
% \And
% Coauthor \\
% Affiliation \\
% Address \\
% \texttt{email} \\
}
\begin{document}
\maketitle
\input{sections/abstract}
\input{sections/intro}
\input{sections/background}
\input{sections/model}
\input{sections/application}
\input{sections/experiment}
\input{sections/related_work}
\input{sections/conclusion}
\input{sections/acknowledgement}
\bibliography{main}
\bibliographystyle{icml}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% \input{sections/checklist}
\input{sections/appendix}
\end{document}
%\newcommand*\modelname{Kaleidoscope}
\newcommand*\modelname{Kaleido}
\NewDocumentCommand{\ying}{ mO{} }{\textcolor{teal}{\textsuperscript{\textit{Ying}}\textsf{\textbf{\small[#1]}}}}
\NewDocumentCommand{\jg}{ mO{} }{\textcolor{red}{\textsuperscript{\textit{Jiatao}}\textsf{\textbf{\small[#1]}}}}
\NewDocumentCommand{\yz}{ mO{} }{\textcolor{blue}{\textsuperscript{\textit{Yizhe}}\textsf{\textbf{\small[#1]}}}}
%%%%% NEW MATH DEFINITIONS %%%%%
\usepackage{amsmath,amsfonts,bm}
% Mark sections of captions for referring to divisions of figures
\newcommand{\figleft}{{\em (Left)}}
\newcommand{\figcenter}{{\em (Center)}}
\newcommand{\figright}{{\em (Right)}}
\newcommand{\figtop}{{\em (Top)}}
\newcommand{\figbottom}{{\em (Bottom)}}
\newcommand{\captiona}{{\em (a)}}
\newcommand{\captionb}{{\em (b)}}
\newcommand{\captionc}{{\em (c)}}
\newcommand{\captiond}{{\em (d)}}
% Highlight a newly defined term
\newcommand{\newterm}[1]{{\bf #1}}
% Figure reference, lower-case.
\def\figref#1{figure~\ref{#1}}
% Figure reference, capital. For start of sentence
\def\Figref#1{Figure~\ref{#1}}
\def\twofigref#1#2{figures \ref{#1} and \ref{#2}}
\def\quadfigref#1#2#3#4{figures \ref{#1}, \ref{#2}, \ref{#3} and \ref{#4}}
% Section reference, lower-case.
\def\secref#1{section~\ref{#1}}
% Section reference, capital.
\def\Secref#1{Section~\ref{#1}}
% Reference to two sections.
\def\twosecrefs#1#2{sections \ref{#1} and \ref{#2}}
% Reference to three sections.
\def\secrefs#1#2#3{sections \ref{#1}, \ref{#2} and \ref{#3}}
% Reference to an equation, lower-case.
\def\eqref#1{equation~\ref{#1}}
% Reference to an equation, upper case
\def\Eqref#1{Equation~\ref{#1}}
% A raw reference to an equation---avoid using if possible
\def\plaineqref#1{\ref{#1}}
% Reference to a chapter, lower-case.
\def\chapref#1{chapter~\ref{#1}}
% Reference to an equation, upper case.
\def\Chapref#1{Chapter~\ref{#1}}
% Reference to a range of chapters
\def\rangechapref#1#2{chapters\ref{#1}--\ref{#2}}
% Reference to an algorithm, lower-case.
\def\algref#1{algorithm~\ref{#1}}
% Reference to an algorithm, upper case.
\def\Algref#1{Algorithm~\ref{#1}}
\def\twoalgref#1#2{algorithms \ref{#1} and \ref{#2}}
\def\Twoalgref#1#2{Algorithms \ref{#1} and \ref{#2}}
% Reference to a part, lower case
\def\partref#1{part~\ref{#1}}
% Reference to a part, upper case
\def\Partref#1{Part~\ref{#1}}
\def\twopartref#1#2{parts \ref{#1} and \ref{#2}}
\def\ceil#1{\lceil #1 \rceil}
\def\floor#1{\lfloor #1 \rfloor}
\def\1{\bm{1}}
\newcommand{\train}{\mathcal{D}}
\newcommand{\valid}{\mathcal{D_{\mathrm{valid}}}}
\newcommand{\test}{\mathcal{D_{\mathrm{test}}}}
\def\eps{{\epsilon}}
% Random variables
\def\reta{{\textnormal{$\eta$}}}
\def\ra{{\textnormal{a}}}
\def\rb{{\textnormal{b}}}
\def\rc{{\textnormal{c}}}
\def\rd{{\textnormal{d}}}
\def\re{{\textnormal{e}}}
\def\rf{{\textnormal{f}}}
\def\rg{{\textnormal{g}}}
\def\rh{{\textnormal{h}}}
\def\ri{{\textnormal{i}}}
\def\rj{{\textnormal{j}}}
\def\rk{{\textnormal{k}}}
\def\rl{{\textnormal{l}}}
% rm is already a command, just don't name any random variables m
\def\rn{{\textnormal{n}}}
\def\ro{{\textnormal{o}}}
\def\rp{{\textnormal{p}}}
\def\rq{{\textnormal{q}}}
\def\rr{{\textnormal{r}}}
\def\rs{{\textnormal{s}}}
\def\rt{{\textnormal{t}}}
\def\ru{{\textnormal{u}}}
\def\rv{{\textnormal{v}}}
\def\rw{{\textnormal{w}}}
\def\rx{{\textnormal{x}}}
\def\ry{{\textnormal{y}}}
\def\rz{{\textnormal{z}}}
% Random vectors
\def\rvepsilon{{\mathbf{\epsilon}}}
\def\rvtheta{{\mathbf{\theta}}}
\def\rva{{\mathbf{a}}}
\def\rvb{{\mathbf{b}}}
\def\rvc{{\mathbf{c}}}
\def\rvd{{\mathbf{d}}}
\def\rve{{\mathbf{e}}}
\def\rvf{{\mathbf{f}}}
\def\rvg{{\mathbf{g}}}
\def\rvh{{\mathbf{h}}}
\def\rvu{{\mathbf{i}}}
\def\rvj{{\mathbf{j}}}
\def\rvk{{\mathbf{k}}}
\def\rvl{{\mathbf{l}}}
\def\rvm{{\mathbf{m}}}
\def\rvn{{\mathbf{n}}}
\def\rvo{{\mathbf{o}}}
\def\rvp{{\mathbf{p}}}
\def\rvq{{\mathbf{q}}}
\def\rvr{{\mathbf{r}}}
\def\rvs{{\mathbf{s}}}
\def\rvt{{\mathbf{t}}}
\def\rvu{{\mathbf{u}}}
\def\rvv{{\mathbf{v}}}
\def\rvw{{\mathbf{w}}}
\def\rvx{{\mathbf{x}}}
\def\rvy{{\mathbf{y}}}
\def\rvz{{\mathbf{z}}}
% Elements of random vectors
\def\erva{{\textnormal{a}}}
\def\ervb{{\textnormal{b}}}
\def\ervc{{\textnormal{c}}}
\def\ervd{{\textnormal{d}}}
\def\erve{{\textnormal{e}}}
\def\ervf{{\textnormal{f}}}
\def\ervg{{\textnormal{g}}}
\def\ervh{{\textnormal{h}}}
\def\ervi{{\textnormal{i}}}
\def\ervj{{\textnormal{j}}}
\def\ervk{{\textnormal{k}}}
\def\ervl{{\textnormal{l}}}
\def\ervm{{\textnormal{m}}}
\def\ervn{{\textnormal{n}}}
\def\ervo{{\textnormal{o}}}
\def\ervp{{\textnormal{p}}}
\def\ervq{{\textnormal{q}}}
\def\ervr{{\textnormal{r}}}
\def\ervs{{\textnormal{s}}}
\def\ervt{{\textnormal{t}}}
\def\ervu{{\textnormal{u}}}
\def\ervv{{\textnormal{v}}}
\def\ervw{{\textnormal{w}}}
\def\ervx{{\textnormal{x}}}
\def\ervy{{\textnormal{y}}}
\def\ervz{{\textnormal{z}}}
% Random matrices
\def\rmA{{\mathbf{A}}}
\def\rmB{{\mathbf{B}}}
\def\rmC{{\mathbf{C}}}
\def\rmD{{\mathbf{D}}}
\def\rmE{{\mathbf{E}}}
\def\rmF{{\mathbf{F}}}
\def\rmG{{\mathbf{G}}}
\def\rmH{{\mathbf{H}}}
\def\rmI{{\mathbf{I}}}
\def\rmJ{{\mathbf{J}}}
\def\rmK{{\mathbf{K}}}
\def\rmL{{\mathbf{L}}}
\def\rmM{{\mathbf{M}}}
\def\rmN{{\mathbf{N}}}
\def\rmO{{\mathbf{O}}}
\def\rmP{{\mathbf{P}}}
\def\rmQ{{\mathbf{Q}}}
\def\rmR{{\mathbf{R}}}
\def\rmS{{\mathbf{S}}}
\def\rmT{{\mathbf{T}}}
\def\rmU{{\mathbf{U}}}
\def\rmV{{\mathbf{V}}}
\def\rmW{{\mathbf{W}}}
\def\rmX{{\mathbf{X}}}
\def\rmY{{\mathbf{Y}}}
\def\rmZ{{\mathbf{Z}}}
% Elements of random matrices
\def\ermA{{\textnormal{A}}}
\def\ermB{{\textnormal{B}}}
\def\ermC{{\textnormal{C}}}
\def\ermD{{\textnormal{D}}}
\def\ermE{{\textnormal{E}}}
\def\ermF{{\textnormal{F}}}
\def\ermG{{\textnormal{G}}}
\def\ermH{{\textnormal{H}}}
\def\ermI{{\textnormal{I}}}
\def\ermJ{{\textnormal{J}}}
\def\ermK{{\textnormal{K}}}
\def\ermL{{\textnormal{L}}}
\def\ermM{{\textnormal{M}}}
\def\ermN{{\textnormal{N}}}
\def\ermO{{\textnormal{O}}}
\def\ermP{{\textnormal{P}}}
\def\ermQ{{\textnormal{Q}}}
\def\ermR{{\textnormal{R}}}
\def\ermS{{\textnormal{S}}}
\def\ermT{{\textnormal{T}}}
\def\ermU{{\textnormal{U}}}
\def\ermV{{\textnormal{V}}}
\def\ermW{{\textnormal{W}}}
\def\ermX{{\textnormal{X}}}
\def\ermY{{\textnormal{Y}}}
\def\ermZ{{\textnormal{Z}}}
% Vectors
\def\vzero{{\bm{0}}}
\def\vone{{\bm{1}}}
\def\vmu{{\bm{\mu}}}
\def\vtheta{{\bm{\theta}}}
\def\va{{\bm{a}}}
\def\vb{{\bm{b}}}
\def\vc{{\bm{c}}}
\def\vd{{\bm{d}}}
\def\ve{{\bm{e}}}
\def\vf{{\bm{f}}}
\def\vg{{\bm{g}}}
\def\vh{{\bm{h}}}
\def\vi{{\bm{i}}}
\def\vj{{\bm{j}}}
\def\vk{{\bm{k}}}
\def\vl{{\bm{l}}}
\def\vm{{\bm{m}}}
\def\vn{{\bm{n}}}
\def\vo{{\bm{o}}}
\def\vp{{\bm{p}}}
\def\vq{{\bm{q}}}
\def\vr{{\bm{r}}}
\def\vs{{\bm{s}}}
\def\vt{{\bm{t}}}
\def\vu{{\bm{u}}}
\def\vv{{\bm{v}}}
\def\vw{{\bm{w}}}
\def\vx{{\bm{x}}}
\def\vy{{\bm{y}}}
\def\vz{{\bm{z}}}
\def\vepsilon{{\bm{\epsilon}}}
\def\vgamma{{\bm{\gamma}}}
% Elements of vectors
\def\evalpha{{\alpha}}
\def\evbeta{{\beta}}
\def\evepsilon{{\epsilon}}
\def\evlambda{{\lambda}}
\def\evomega{{\omega}}
\def\evmu{{\mu}}
\def\evpsi{{\psi}}
\def\evsigma{{\sigma}}
\def\evtheta{{\theta}}
\def\eva{{a}}
\def\evb{{b}}
\def\evc{{c}}
\def\evd{{d}}
\def\eve{{e}}
\def\evf{{f}}
\def\evg{{g}}
\def\evh{{h}}
\def\evi{{i}}
\def\evj{{j}}
\def\evk{{k}}
\def\evl{{l}}
\def\evm{{m}}
\def\evn{{n}}
\def\evo{{o}}
\def\evp{{p}}
\def\evq{{q}}
\def\evr{{r}}
\def\evs{{s}}
\def\evt{{t}}
\def\evu{{u}}
\def\evv{{v}}
\def\evw{{w}}
\def\evx{{x}}
\def\evy{{y}}
\def\evz{{z}}
% Matrix
\def\mA{{\bm{A}}}
\def\mB{{\bm{B}}}
\def\mC{{\bm{C}}}
\def\mD{{\bm{D}}}
\def\mE{{\bm{E}}}
\def\mF{{\bm{F}}}
\def\mG{{\bm{G}}}
\def\mH{{\bm{H}}}
\def\mI{{\bm{I}}}
\def\mJ{{\bm{J}}}
\def\mK{{\bm{K}}}
\def\mL{{\bm{L}}}
\def\mM{{\bm{M}}}
\def\mN{{\bm{N}}}
\def\mO{{\bm{O}}}
\def\mP{{\bm{P}}}
\def\mQ{{\bm{Q}}}
\def\mR{{\bm{R}}}
\def\mS{{\bm{S}}}
\def\mT{{\bm{T}}}
\def\mU{{\bm{U}}}
\def\mV{{\bm{V}}}
\def\mW{{\bm{W}}}
\def\mX{{\bm{X}}}
\def\mY{{\bm{Y}}}
\def\mZ{{\bm{Z}}}
\def\mBeta{{\bm{\beta}}}
\def\mPhi{{\bm{\Phi}}}
\def\mLambda{{\bm{\Lambda}}}
\def\mSigma{{\bm{\Sigma}}}
% Tensor
\DeclareMathAlphabet{\mathsfit}{\encodingdefault}{\sfdefault}{m}{sl}
\SetMathAlphabet{\mathsfit}{bold}{\encodingdefault}{\sfdefault}{bx}{n}
\newcommand{\tens}[1]{\bm{\mathsfit{#1}}}
\def\tA{{\tens{A}}}
\def\tB{{\tens{B}}}
\def\tC{{\tens{C}}}
\def\tD{{\tens{D}}}
\def\tE{{\tens{E}}}
\def\tF{{\tens{F}}}
\def\tG{{\tens{G}}}
\def\tH{{\tens{H}}}
\def\tI{{\tens{I}}}
\def\tJ{{\tens{J}}}
\def\tK{{\tens{K}}}
\def\tL{{\tens{L}}}
\def\tM{{\tens{M}}}
\def\tN{{\tens{N}}}
\def\tO{{\tens{O}}}
\def\tP{{\tens{P}}}
\def\tQ{{\tens{Q}}}
\def\tR{{\tens{R}}}
\def\tS{{\tens{S}}}
\def\tT{{\tens{T}}}
\def\tU{{\tens{U}}}
\def\tV{{\tens{V}}}
\def\tW{{\tens{W}}}
\def\tX{{\tens{X}}}
\def\tY{{\tens{Y}}}
\def\tZ{{\tens{Z}}}
% Graph
\def\gA{{\mathcal{A}}}
\def\gB{{\mathcal{B}}}
\def\gC{{\mathcal{C}}}
\def\gD{{\mathcal{D}}}
\def\gE{{\mathcal{E}}}
\def\gF{{\mathcal{F}}}
\def\gG{{\mathcal{G}}}
\def\gH{{\mathcal{H}}}
\def\gI{{\mathcal{I}}}
\def\gJ{{\mathcal{J}}}
\def\gK{{\mathcal{K}}}
\def\gL{{\mathcal{L}}}
\def\gM{{\mathcal{M}}}
\def\gN{{\mathcal{N}}}
\def\gO{{\mathcal{O}}}
\def\gP{{\mathcal{P}}}
\def\gQ{{\mathcal{Q}}}
\def\gR{{\mathcal{R}}}
\def\gS{{\mathcal{S}}}
\def\gT{{\mathcal{T}}}
\def\gU{{\mathcal{U}}}
\def\gV{{\mathcal{V}}}
\def\gW{{\mathcal{W}}}
\def\gX{{\mathcal{X}}}
\def\gY{{\mathcal{Y}}}
\def\gZ{{\mathcal{Z}}}
% Sets
\def\sA{{\mathbb{A}}}
\def\sB{{\mathbb{B}}}
\def\sC{{\mathbb{C}}}
\def\sD{{\mathbb{D}}}
% Don't use a set called E, because this would be the same as our symbol
% for expectation.
\def\sF{{\mathbb{F}}}
\def\sG{{\mathbb{G}}}
\def\sH{{\mathbb{H}}}
\def\sI{{\mathbb{I}}}
\def\sJ{{\mathbb{J}}}
\def\sK{{\mathbb{K}}}
\def\sL{{\mathbb{L}}}
\def\sM{{\mathbb{M}}}
\def\sN{{\mathbb{N}}}
\def\sO{{\mathbb{O}}}
\def\sP{{\mathbb{P}}}
\def\sQ{{\mathbb{Q}}}
\def\sR{{\mathbb{R}}}
\def\sS{{\mathbb{S}}}
\def\sT{{\mathbb{T}}}
\def\sU{{\mathbb{U}}}
\def\sV{{\mathbb{V}}}
\def\sW{{\mathbb{W}}}
\def\sX{{\mathbb{X}}}
\def\sY{{\mathbb{Y}}}
\def\sZ{{\mathbb{Z}}}
% Entries of a matrix
\def\emLambda{{\Lambda}}
\def\emA{{A}}
\def\emB{{B}}
\def\emC{{C}}
\def\emD{{D}}
\def\emE{{E}}
\def\emF{{F}}
\def\emG{{G}}
\def\emH{{H}}
\def\emI{{I}}
\def\emJ{{J}}
\def\emK{{K}}
\def\emL{{L}}
\def\emM{{M}}
\def\emN{{N}}
\def\emO{{O}}
\def\emP{{P}}
\def\emQ{{Q}}
\def\emR{{R}}
\def\emS{{S}}
\def\emT{{T}}
\def\emU{{U}}
\def\emV{{V}}
\def\emW{{W}}
\def\emX{{X}}
\def\emY{{Y}}
\def\emZ{{Z}}
\def\emSigma{{\Sigma}}
% entries of a tensor
% Same font as tensor, without \bm wrapper
\newcommand{\etens}[1]{\mathsfit{#1}}
\def\etLambda{{\etens{\Lambda}}}
\def\etA{{\etens{A}}}
\def\etB{{\etens{B}}}
\def\etC{{\etens{C}}}
\def\etD{{\etens{D}}}
\def\etE{{\etens{E}}}
\def\etF{{\etens{F}}}
\def\etG{{\etens{G}}}
\def\etH{{\etens{H}}}
\def\etI{{\etens{I}}}
\def\etJ{{\etens{J}}}
\def\etK{{\etens{K}}}
\def\etL{{\etens{L}}}
\def\etM{{\etens{M}}}
\def\etN{{\etens{N}}}
\def\etO{{\etens{O}}}
\def\etP{{\etens{P}}}
\def\etQ{{\etens{Q}}}
\def\etR{{\etens{R}}}
\def\etS{{\etens{S}}}
\def\etT{{\etens{T}}}
\def\etU{{\etens{U}}}
\def\etV{{\etens{V}}}
\def\etW{{\etens{W}}}
\def\etX{{\etens{X}}}
\def\etY{{\etens{Y}}}
\def\etZ{{\etens{Z}}}
% The true underlying data generating distribution
\newcommand{\pdata}{p_{\rm{data}}}
% The empirical distribution defined by the training set
\newcommand{\ptrain}{\hat{p}_{\rm{data}}}
\newcommand{\Ptrain}{\hat{P}_{\rm{data}}}
% The model distribution
\newcommand{\pmodel}{p_{\rm{model}}}
\newcommand{\Pmodel}{P_{\rm{model}}}
\newcommand{\ptildemodel}{\tilde{p}_{\rm{model}}}
% Stochastic autoencoder distributions
\newcommand{\pencode}{p_{\rm{encoder}}}
\newcommand{\pdecode}{p_{\rm{decoder}}}
\newcommand{\precons}{p_{\rm{reconstruct}}}
\newcommand{\laplace}{\mathrm{Laplace}} % Laplace distribution
\newcommand{\E}{\mathbb{E}}
\newcommand{\Ls}{\mathcal{L}}
\newcommand{\R}{\mathbb{R}}
\newcommand{\emp}{\tilde{p}}
\newcommand{\lr}{\alpha}
\newcommand{\reg}{\lambda}
\newcommand{\rect}{\mathrm{rectifier}}
\newcommand{\softmax}{\mathrm{softmax}}
\newcommand{\sigmoid}{\sigma}
\newcommand{\softplus}{\zeta}
\newcommand{\KL}{D_{\mathrm{KL}}}
\newcommand{\Var}{\mathrm{Var}}
\newcommand{\standarderror}{\mathrm{SE}}
\newcommand{\Cov}{\mathrm{Cov}}
% Wolfram Mathworld says $L^2$ is for function spaces and $\ell^2$ is for vectors
% But then they seem to use $L^2$ for vectors throughout the site, and so does
% wikipedia.
\newcommand{\normlzero}{L^0}
\newcommand{\normlone}{L^1}
\newcommand{\normltwo}{L^2}
\newcommand{\normlp}{L^p}
\newcommand{\normmax}{L^\infty}
\newcommand{\parents}{Pa} % See usage in notation.tex. Chosen to match Daphne's book.
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator{\sign}{sign}
\DeclareMathOperator{\Tr}{Tr}
\let\ab\allowbreak
\begin{figure}[h]
\centering
\vspace{-20pt}
\includegraphics[width=\textwidth]{figures/teaserv3.jpeg}
\caption{\footnotesize Text conditioned high-resolution samples of variable aspect ratios generated from our $3.8$B {\model} model. Resolutions are adjusted for the ease of visualization.}
\label{fig:teaser}
\end{figure}
\begin{abstract}
We present \model{}, a scalable generative model based on normalizing flows that achieves strong performance on high-resolution image synthesis.
\model{}'s main building block is Transformer Autoregressive Flow (TARFlow), which combines normalizing flows with Autoregressive Transformer architectures and has recently achieved impressive results in image modeling. In this work, we first establish the theoretical universality of TARFlow for modeling continuous distributions. Building on this foundation, we introduce a set of architectural and algorithmic innovations that significantly enhance the scalability: (1) a deep-shallow design where a deep Transformer block captures most of the model’s capacity, followed by a few shallow Transformer blocks that are computationally cheap yet contribute non-negligibly, (2) learning in the latent space of pretrained autoencoders, which proves far more effective than modeling pixels directly, and (3) a novel guidance algorithm that substantially improves sample quality. Crucially, our model remains a single, end-to-end normalizing flow, allowing exact maximum likelihood training in continuous space without discretization. \model{} achieves competitive results in both class- and text-conditional image generation, with sample quality approaching that of state-of-the-art diffusion models. To our knowledge, this is the \textbf{first} successful demonstration of normalizing flows at this scale and resolution.
\end{abstract}
\section{Introduction}
Recent years have witnessed remarkable progress in high-resolution text-to-image generative modeling, with state-of-the-art approaches predominantly falling into two distinct categories. On one hand, diffusion models~\citep{ho2020denoising,rombach2022high,peebles2023scalable,esser2024scaling} operating in continuous space have set new benchmarks in image quality. However, their reliance on iterative denoising processes renders both training and inference computationally intensive. On the other hand, autoregressive image generation methods~\citep{yu2022scaling,sun2024autoregressive,tian2024visual}—inspired by the success of large language models~\citep[LLMs,][]{gpt3,dubey2024llama}—avoid such inefficiencies by modeling images in discrete space via quantization; yet, this quantization can impose stringent limitations and adversely affect fidelity. More recently, a promising trend has emerged to explore hybrid models~\citep{li2024autoregressive,gu2024dart,fan2024fluid} that apply autoregressive techniques directly in continuous space. However, the inherently distinct characteristics of these two paradigms introduce additional complexity in effective unification.
In this paper, we turn our eyes on the yet another modeling approach -- Normalizing Flows~\citep[NFs,][]{pmlr-v37-rezende15,dinh2016density}, a family of likelihood based models that have received relatively little attention in the recent wave of Generative AI. We start from inspecting TARFlow~\citep[][]{zhai2024normalizing}, a recently proposed model that combines a powerful Transformer architecture with autoregressive flows~\citep[AFs,][]{kingma2016improved, papamakarios2017masked}. While TARFlow demonstrates promising results on the potential of NFs as a modeling principle, it remains unclear whether it can perform as a scalable method, in comparison to other approaches such as diffusion and discrete autoregressive models.
To this end, we propose \model{}, a family of generative models that shows for the \textbf{first-time} that NF models can successfully generalize to high-resolution and large-scale image modeling.
We first provide a theoretical insight on \emph{why AFs can be capable generative models} by showing the universality of multi-block AFs in modeling continuous distributions. On top of this, we propose a novel \emph{deep–shallow} architecture. We found that the architecture configuration, e.g., the number of flows as well as the depth and width of the Transformer for each flow, plays a pivotal role to the model's performance. While TARFlow~\citep{zhai2024normalizing} proposes to uniformly allocate model depth among all flows, we found that it is beneficial to have a skewed architecture design, where we allocate most of the model parameters to the first AF block (i.e., the one closest to the prior), which is followed by a few shallow but non-negligible blocks.
Importantly, our model still yields a stand-alone normalizing-flow framework that supports end-to-end maximum-likelihood training in continuous space, thereby sidestepping the quantization limits inherent to discrete models. Rather than operating directly in data space, we instead learn AFs in the latent space of pretrained autoencoders. Crucially, we demonstrate that NFs align naturally with compressed latents—an intuitive yet vital observation—enabling far superior modeling of high-resolution inputs, as verified in our experiments, compared with training directly on pixels. Similar to TARFlow, noise injection proves essential: by fine-tuning the decoder, we train the model on noisy latents and at the same time simplify the original sampling pipeline.
Moreover, we revisit the classifier-free guidance (CFG) algorithm for AFs from a more principled way and propose a novel guidance algorithm, which substantially improves image quality, especially at high guidance weights in text-to-image generation tasks.
Together, these innovations represent the first demonstration of NF models applied to large-scale, high-resolution image generation. Our approach offers a scalable and efficient alternative to conventional diffusion-based and autoregressive approaches, achieving competitive performance on benchmarks for both class-conditioned image and large-scale text-to-image synthesis.
Moreover, our framework is highly flexible, and we demonstrate that it easily enables interesting settings such as image inpainting and instruction based image editing by finetuning.
\section{Preliminaries}
\subsection{Normalizing Flows}
In this paper, we consider Normalizing Flows~\citep[NFs,][]{pmlr-v37-rezende15,dinh2014nice,dinh2016density} as the class of likelihood method that follows the change of variable formula. Given continuous inputs $\vx \sim \pdata$, $\vx \in \mathbb{R}^{D}$, a NF learns an invertible transformation $f_\theta: \mathbb{R}^{D} \mapsto \mathbb{R}^{D}$ (with $\theta$ being the parameters) which maps data $\vx$ into the noise space $f_\theta(\vx)$, and can be trained with maximum likelihood estimation (MLE):
\begin{equation}
\label{eq:nf_loss}
\max_\theta~ \mathbb{E}_{\vx\sim \pdata}\log \pnf(\vx;\theta) = \log p_{0}(f_\theta(\vx);\theta) + \log\left(\left\lvert\det\left(\frac{\partial f_\theta(\vx)}{\partial\vx}\right)\right\rvert\right),
\end{equation}
where the first term rewards sending data to high-density regions of the prior \(p_0\), while the Jacobian term penalizes excessive local volume shrinkage, ensuring the transformation remains bijective and does not collapse nearby points onto a lower-dimensional set.
One automatically obtains a generative model by inverting $f_\theta$, with a sampling procedure $\vz \sim p_{0}(\vz)$, $\vx = f_\theta^{-1}(\vz)$.
\subsection{Autoregressive Flows and TARFlow}
\label{sec.af}
An interesting variant of NFs is autoregressive flows~\citep[AFs,][]{kingma2016improved,papamakarios2017masked}.
In the simplest affine form, an AF constructs $\vz = f_\theta(\vx) = \{\mu_\theta, \sigma_\theta\}(\vx)$ as a standalone invertible model with the forward ($\vx\rightarrow\vz$) and sampling ($\vz\rightarrow\vx$) process:
\begin{equation}
\vz_d = \left(\vx_d - \mu_\theta(\vx_{<d})\right) / \sigma_\theta(\vx_{<d}), ~~~~~~
\vx_d = \mu_\theta(\vx_{<d}) + \sigma_\theta(\vx_{<d}) \cdot \vz_d,~~ \forall d \in [1, D],
\label{eq.af_step}
\end{equation}
where $\vx_0$ is a constant \texttt{<sos>}. This can be seen as ``next-token prediction'' with affine transformation, and training with \cref{eq:nf_loss} where the Jacobian term becomes extremely simple as $-\!\sum_{d=1}^D \log\sigma_\theta(\vx_{<d}).$
The extension to multi-channel inputs \(\vx\in\mathbb{R}^{D\times C}\) (e.g., $C\!=\!3$ for RGB image) is immediate as channels at each step can be treated as conditionally independent. We \emph{omit} the channel dim henceforth.
Recently, \citet{zhai2024normalizing} introduced TARFlow, a compelling framework for building performant NFs for image data. Specifically, TARFlows can be viewed as a special form of AFs by pairing causal-Transformer blocks with an extension of classical AF formulation -- stacking multiple AF layers whose autoregressive ordering \textbf{alternates} from one layer to the next.
To be concrete, with $T$ flows, we have
$
\vz = f^T_\theta \circ f^{2}_\theta \circ \cdots \circ f^1_\theta(\vx),
$
where each block $f^t_\theta(.)$ processes the input in its own ordering $\vx_\pi = (\vx_{\pi_1},\ldots, \vx_{\pi_D})$ (a permutation of $\{\vx_1\ldots \vx_D\}$), enabling the stack to capture dependencies in both directions of the data sequence. Training is still performed end-to-end:
\begin{equation}
\max_\theta~ \mathbb{E}_{\vx\sim\pdata} \log \paf(\vx;\theta) = -\frac{1}{2}\|\vz\|^2_2 - \sum_{t=1}^T\sum_{d=1}^{D} \log \sigma^t_\theta(\vx^{t}_{\pi_{<d}}),
\label{eq:tarflow_loss}
\end{equation}
where $\vx^{t} = f^t_\theta (\vx^{t-1})$ defines the forward propagation (\cref{eq.af_step});
we denote the data $\vx=\vx^0$ and the final output $\vz=\vx^T$ is modeled with standard Gaussian.
Additionally, \citet{zhai2024normalizing} also proposed several techniques to improve the modeling capability, including noise augmented training, score-based denoising and incorporating guidance~\citep{ho2021classifier}.
\section{\model}
In this section, we propose \underline{S}calable \underline{T}ransformer \underline{A}uto\underline{r}egressive \underline{Flow} (\model{}), a method that pushes the frontier of NF based high-resolution image generation.
We first establish—on theoretical grounds—AFs' expressivity as a general modeling method in \cref{sec.why_af}, based on which we propose our core approaches by improving TARFlow in several key aspects: (1) a better architecture configuration (\cref{sec.arch}), (2) a working recipe of learning in the latent space (\cref{sec.latent}) and (3) a novel guidance algorithm (\cref{sec.cfg}).
An illustration of the learning and inference pipeline is presented in \cref{fig:pipeline}.
\input{sections/model_why_af_work}
\input{sections/model_architecture}
\input{sections/model_latent}
\input{sections/model_guidance}
\subsection{Why TARFlows are Capable Generative Models?}
\label{sec.why_af}
While empirical results confirm that TARFlow is highly competitive~\citep{zhai2024normalizing}, we ask—from a modeling perspective—whether they are expressive enough to warrant scaling. Here, we claim:
\begin{comment}
\begin{remarkbox}
\textbf{Proposition~1.}~ Autoregressive flows with $T \geq 2$ blocks of $D$ steps and alternating generation orderings are universal approximators of continuous densities on $\mathbb{R}^{D}$ except for the last dimension.
\end{remarkbox}
\emph{Sketch of Proof.}~ Consider $T=2$. Without loss of generality, we model $f_\theta = f^1_\theta\circ f^2_\theta$ where $f^{1}_\theta$ and $f^{2}_\theta$ employ reversed orderings (forward and backward) for data $\vx \in \mathbb{R}^{D}$ (see \jg{todo}):
\begin{equation}
\vx_d = \mu^2_\theta(\vx_{<d}) + \sigma^2_\theta(\vx_{<d}) \cdot \vy_d, ~~
\vy_d = \mu^1_\theta(\vy_{>d}) + \sigma^1_\theta(\vy_{>d}) \cdot \vz_d, ~~
\vz_d \sim \mathcal{N}(\mathbf{0}, I),~~ d \in [1, D],
\label{eq.2_af}
\end{equation}
%where $\pi_{<d}$ represents the ordering indices $\vy_d$ is conditioned on from $f^2_\theta$.
This yields the autoregressive factorization $p(\vx)=\prod_{d=1}^{D} p(\vx_{d}\mid\vx_{<d})$ as follows:
\begin{equation}
p(\vx_{d}\mid\vx_{<d})=\int \mathcal{N}\bigl(\vx_{d}\mid\hat\mu(\vx_{<d}, \vz_{>d}),\hat\sigma^{2}(\vx_{<d}, \vz_{>d})I\bigr)\cdot
p(\vz_{>d}\mid\vx_{<d}) \mathrm{d}\vz_{>d},
\label{eq.af_cond}
\end{equation}
where $\hat{\mu}$, $\hat{\sigma}$ are computed affine coefficients from \cref{eq.2_af}.
Each conditional probability in \cref{eq.af_cond} is an \emph{infinite Gaussian mixture} as long as $\vz_{>d} \neq \emptyset$, %\footnote{Otherwise, such probability degenerates to a single-mode Gaussian.},
a family that is dense in the space of continuous densities. When $d=D$, \cref{eq.2_af} degenerates to single-mode Gaussian.
Hence $f$ can approximate any distribution for $\vx$ arbitrarily well except for the last dimension where AF can only model it in uni-modal. The universality follows for all $T>2$.
More details including degenerated boundary cases in .\hfill$\square$
\end{comment}
%alternative version TC
\begin{remarkbox}\label{prop:expressive}
Stacked autoregressive flows with $T \geq 3$ blocks of $D$ autoregressive steps and alternating orderings are universal approximators for any continuous density $p\in L^1(\mathbb{R}^{D})$. %$\mathcal{P}=\left\{p\big| p\in L^1(\mathbb{R}^{D-1}),p(\vx)\geq 0, \int_{\mathbb{R}^{D-1}}p(\vx)d\vx=1\right\}$.
\end{remarkbox}
%Here we denote $\vx_{<d}:=\left\{\vx_i\big|1\leq i<d\leq D\right\}$.
%\begin{comment}
\emph{Sketch of Proof.}~
First consider $T=2$. Without loss of generality, we model $f_\theta = f^a_\theta\circ f^b_\theta$ where $f^{a}_\theta$ and $f^{b}_\theta$ employ reversed orderings (forward and backward) for data $\vx \in \mathbb{R}^{D}$ (see~\cref{fig:2block_af}):
\begin{equation}
\vx_d = \mu^b_\theta(\vx_{<d}) + \sigma^b_\theta(\vx_{<d}) \cdot \vy_d, ~~
\vy_d = \mu^a_\theta(\vy_{>d}) + \sigma^a_\theta(\vy_{>d}) \cdot \vz_d, ~~
\vz_d \sim \mathcal{N}(\mathbf{0}, I),~~ d \in [1, D],
\label{eq.2_af}
\end{equation}
%where $\pi_{<d}$ represents the ordering indices $\vy_d$ is conditioned on from $f^2_\theta$.
This yields the autoregressive factorization $p(\vx)=\prod_{d=1}^{D} p(\vx_{d}\mid\vx_{<d})$ as follows:
\begin{equation}
p(\vx_{d}\mid\vx_{<d})=\int \mathcal{N}\bigl(\vx_{d}\mid\hat\mu_\theta(\vx_{<d}, \vy_{>d}),\hat\sigma_\theta^{2}(\vx_{<d}, \vy_{>d})I\bigr)\cdot
p(\vy_{>d}\mid\vx_{<d}) \mathrm{d}\vy_{>d},
\label{eq.af_cond}
\end{equation}
where $\hat\mu_\theta = \mu^b_\theta(\vx_{<d})+ \mu^a_\theta(\vy_{>d})\sigma^b_\theta(\vx_{<d})$, $\hat\sigma_\theta = \sigma^a_\theta(\vy_{>d})\sigma^b_\theta(\vx_{<d})$
defined in \cref{eq.2_af}.
For every $d < D$, we have \(\vy_{>d}\neq\varnothing\). The integral in \cref{eq.af_cond} forms an \emph{infinite Gaussian mixture}, a family that is \textbf{dense} in the continuous density based on the universal-approximation theorem of Gaussian mixtures (p.~65 of \citet{Goodfellow-et-al-2016}) with the expressive power of neural networks.
\begin{wrapfigure}{r}{0.3\textwidth}
\centering
\vspace{0pt}
\includegraphics[width=0.3\textwidth]{figures/2block_af.jpeg}
\caption{\footnotesize Ex. of 2-block AFs.}
\label{fig:2block_af}
\vspace{-10pt}
\end{wrapfigure}
For the final coordinate \(d=D\) we have \(\vy_{>D}=\varnothing\). \cref{eq.af_cond} reduces to a single Gaussian and the universality is lost. However, this restriction is lifted by extending additional flows ($T\geq 3$) to re‑introduce latent variables or appending an augmented Gaussian variable.
Additional derivation details appear in the \cref{app:A}.
\hfill$\square$
The preceding proposition clarifies why we can safely scale-up AFs on large data. Even in the minimal setting $T=2$ where full universality is not attained, the resulting limitation is negligible in high‑dimensional domains such as natural images.
\subsection{Proposed Architecture}
\label{sec.arch}
\begin{wrapfigure}{r}{0.4\textwidth}
\centering
\vspace{-9pt}
\includegraphics[width=0.4\textwidth]{figures/guidance_example.jpg}
\caption{\footnotesize Top to bottom, guiding the first 0, 3, 8 flow blocks with a TARFlow model with 8 flow blocks. We see that guidance is only effective up to the top 3 blocks.}
\label{fig:guide}
\vspace{-10pt}
\end{wrapfigure}
The derivation in \cref{sec.why_af} motivates a redesign of scalable AF architectures within realistic computational budgets, emphasizing that we need not greatly expand the number of flow blocks—indeed (even $T=2$ often suffices). However, the remark leaves unresolved how best to allocate compute across those blocks. We first inspect the proposed architecture configuration in TARFlow, which suggests to allocate equal sized Transformer layers for each flow. Interestingly, in our reproduced TARFlow results, we see that most effective compute (measured through the lens of guidance) concentrates in just the top few AF blocks (see motivating examples \cref{fig:guide}). We conjecture that end‑to‑end training drives the network to exploit layers closest to the noise, a behavior that contrasts that of diffusion models.
\vspace{-5pt}\paragraph{Deep-shallow Architecture}
Our architecture can be intuitively considered as an extension of standard autoregressive language models (e.g., LLaMA~\citep{dubey2024llama}) with a general \emph{deep-shallow} design. At inference time, a deep AF block first autoregressively generates $\vx^1$ from noise $\vz$, followed by a sequence of shallow AF blocks that iteratively refine it to $\vx^N$, all while keeping the total number of blocks $T$ small. Given a total depth budget $L$, we instantiate the model as $l(T)$: one deep $l$-layer block and $T{-}1$ shallow 2-layer blocks, satisfying $L = l + 2(T{-}1)$.
This asymmetric design turns the deep block into a \emph{Gaussian language model}, while the shallow stack plays the role of a \emph{learned image tokenizer}.
\vspace{-5pt}\paragraph{Conditional \model{}}
This design naturally extends to conditional generation by simply prepending the control signal (e.g., class label, caption) to the input of the flow. Interestingly, our preliminary experiments show that conditioning only the deep block—while leaving the shallow blocks to focus solely on local image refinement—incurs no loss in performance. This not only simplifies the overall architecture, but also enables seamless initialization of the deep block with any pre-trained language model, without major modifications. As a result, our image generator can be directly integrated into any LLM’s semantic space, eliminating the need for a separate text encoder. % ~\citep{gu2024dart}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/pipeline_v2.jpeg}
\caption{\footnotesize An illustration of the {\bf autoregressive} inference (left) and {\bf parallel} training (right) process of our proposed model for text-to-image generation. The upward (green) and downward (purple) arrows refers to the inverse and forward AF step as shown in \cref{eq.af_step}.}
%The deep block can be initialized with any pretrained LLMs.}
\vspace{-10pt}
\label{fig:pipeline}
\end{figure}
\subsection{Moving to Latent Space}
\label{sec.latent}
Analogy to Stable Diffusion~\citep[SD,][]{rombach2022high} w.r.t standard diffusion models, \model{} directly models the latent space of a pretrained autoencoders $\vx\approx \mathcal{D}(\tilde{\vx}), \tilde{\vx} = \mathcal{E}(\vx)$, enabling high-resolution image generation. For instance, when using SD-1.4 autoencoder\footnote{\url{https://huggingface.co/stabilityai/sd-vae-ft-mse}.}, one can reduce input shape from $256\times 256$ to $32\times 32$.
As noted by \citet{zhai2024normalizing}, injecting a proper amount of Gaussian noise, instead of small dequantization noise \citet{dinh2016density,ho2019flow++}, is crucial for stable training and high quality sampling. This then makes it necessary to perform an additional score-based denoising step to clean up the noise components in the samples \citet{zhai2024normalizing}.
In the context of latent normalizing flows, however, the added noise becomes an integral component of the latent representation. Specifically, we encode each sample as
%\begin{equation}
$ \tilde{\vx} \sim \pencoder = \mathcal{N}\left(\mathcal{E}(\vx); \sigma_L^2I\right).$
%\end{equation}
We perform preliminary search for the noise scale ($\sigma_L$) to based on the choice of autoencoders. For example, we set $\sigma_L = 0.3$ throughout the paper.
\vspace{-5pt}\paragraph{Learning}
Learning in the latent space leaves additional flexibility that the flow model can focus on high-level semantics and leave the low-level local details with the pixel decoder.
In this way, AF acts as a learnable prior for the latents. Following VAEs~\citep{kingma2013auto}, we optimize the entire model by maximizing the evidence lower-bound (ELBO) where the entropy term is constant:
\begin{equation}
\max_{\theta,\phi} \mathbb{E}_{\tilde{\vx}\sim \pencoder(\tilde{\vx}|\vx), \vx\sim \pdata}\left[
\log \paf(\tilde{\vx};\theta) + \log \pdecoder(\vx | \tilde{\vx}; \phi)
- \log \pencoder(\tilde{\vx} | \vx)
\right],
\label{eq.training}
\end{equation}
where $\phi$ are the parameters of decoder $\pdecoder$ which transforms the noisy latents back to the pixel space.
Here, we jointly train the AF prior and pixel decoder, freezing the encoder distribution --as in SD--, which stabilizes training and decouples their optimization. Relaxing the encoder $\pencoder$ and training with the full ELBO loss including entropy regularization are left for future work.
\vspace{-5pt}\paragraph{Pixel Decoder}
As shown in \cref{eq.training}, the prertaiend decoder has to be adapted in order to decode from the noisy latents.
Different from \citet{zhai2024normalizing} which relies on gradient-based denoising, modeling in the latent allows a simpler solution by directly fine-tuning the decoder over noisy latents:
\begin{equation}
\min_{\mathcal{\phi}}\mathcal{L}\left(\mathcal{D}(\mathcal{E}(\vx + \sigma\mathbf{\epsilon}); \phi), \vx\right),
\end{equation}
where following \citet{taming}, $\mathcal{L} = \mathcal{L}_{\textrm{L2}} + \mathcal{L}_{\textrm{LPIPS}} + \beta\mathcal{L}_{\textrm{GAN}}$. We empirically observe consistently better performance than score-based denoising technique proposed in \citep{zhai2024normalizing}, with FID decreasing from 2.96 to 2.40 on ImageNet-256. See \cref{app:C} for more discussions.
\subsection{Revisiting Classifier-Free Guidance for Autoregressive Flows}
\label{sec.cfg}
Classifier-free guidance (CFG), originally introduced for diffusion models~\citep{ho2021classifier}, has become a cornerstone in modern generative modeling, proving broadly effective across various architectures, including AR models~\citep{yu2022scaling}. At a high level, CFG amplifies the difference between conditional and unconditional predictions, encouraging more mode-seeking behavior.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/compare_cfg.jpeg}
\caption{\footnotesize (a) Guidance from TARFlow~\citep{zhai2024normalizing} (b) Proposed guidance on ImageNet 256 $\times$ 256.}
\label{fig:compare_cfg}
\vspace{-15pt}
\end{figure}
In the context of AFs, \citet{zhai2024normalizing} made the first attempt to apply CFG by linearly extrapolating the mean and variance at each step (\cref{eq.af_step}): $\tilde{\mu}_c = \mu_c + \omega (\mu_c - \mu_u)$ and $\tilde{\sigma}_c = \sigma_c + \omega (\sigma_c - \sigma_u)$\footnote{We use $c$ and $u$ to denote the conditional and unconditional predictions, respectively.},
where $\omega > 0$ denotes the guidance weight.
While effective to some extent, this naïve formulation lacks principled justification, leaving unclear how $\mu$ and $\sigma$ should be jointly modulated under guidance. Furthermore, as shown in \cref{fig:compare_cfg}, this approach becomes unstable at high guidance weights—precisely the regime required for visually compelling results in text-to-image generation.
We propose to revisit CFG from the perspective of score function, the original intuition of \citet{ho2021classifier}. In short, we want to sample from a guided distribution $\tilde{p}$ which score satisfies:
\begin{equation}
\nabla_\vx \log \tilde{p}_c(\vx) = \nabla_\vx \log p_c(\vx) + \omega \left(\nabla_\vx \log p_c(\vx) - \nabla_\vx\log p_u(\vx) \right).
\label{eq.score_cfg}
\end{equation}
It is generally non-trivial to determine $\tilde{p}_c$ for every flow block. Fortunately, under the design of our proposed model, guidance is only required in the deep block, which functions as a \emph{Gaussian Language Model} (\cref{sec.arch}). Therefore, \cref{eq.score_cfg} can be easily simplified into the following:
\begin{remarkbox}\label{prop:new_cfg}
Given $p_u = \mathcal{N}(\mu_u, \sigma^2_u I)$, and $p_c = \mathcal{N}(\mu_c, \sigma^2_c I)$, the guided distribution $\tilde{p}_c$ is also Gaussian $\tilde{p}_c = \mathcal{N}(\tilde{\mu}_c, \tilde{\sigma}_c^2 I)$ and satisfies:
\begin{equation}
\tilde{\mu}_c = \mu_c + \frac{\omega s}{1+\omega -\omega s}\cdot (\mu_c - \mu_u), ~~~
\tilde{\sigma}_c = \frac{1}{\sqrt{1 + \omega - \omega s}}\cdot \sigma_c,
\label{eq.proposed_cfg}
\end{equation}
where $s = \sigma^2_c / \sigma^2_u$ and $\omega > 0$.
\end{remarkbox}
\emph{proof}:
A detailed derivation is provided in the \cref{app:A}.
\hfill$\square$
Notably, when $\sigma_c = \sigma_u$, \cref{eq.proposed_cfg} reduces to the standard CFG used in diffusion models. However, directly applying \cref{eq.proposed_cfg} can lead to severe numerical instability, as the denominator $1 + \omega - \omega s$ may approach zero or even become negative. To address this, we propose clipping $s$ via $s = \textsc{clip}(s, 0, 1)$, motivated by the intuition that the guided distribution should be more \emph{mode-seeking} than the original, implying that $1 + \omega - \omega s \geq 1$ for any $\omega$, therefore $s \leq 1$.
\subsection{Applications}
\begin{wrapfigure}{r}{0.5\textwidth}
\centering
\vspace{-40pt}
\includegraphics[width=0.5\textwidth]{figures/application_example_merge.jpeg}
\caption{\footnotesize (a) image inpainting (b) interactive editing.}
\label{fig:application_example}
\vspace{-20pt}
\end{wrapfigure}
%\vspace{-0.2cm}
\model{} is a versatile generative model that not only produces diverse, high-quality images under various conditions but also extends naturally to downstream applications. We showcase two examples: image inpainting and editing.
\vspace{-5pt}
\paragraph{Training-Free Inpainting}
We first map the masked image to the latent space, replacing masked regions with Gaussian noise. Reverse sampling is then performed, restoring unmasked pixels with ground truth. We perform generation iteratively until the final inpainted output.
\vspace{-5pt}
\paragraph{Interactive Generation and Editing}
We finetune \model{} on an image editing dataset (\cref{fig:application_example}b), enabling joint modeling of generation and editing with a single conditional AF model. Its invertibility also allows direct image encoding, making it suitable for interactive use.
%\vspace{-0.2cm}
\section{Experiments}
%\vspace{-0.2cm}
\subsection{Experimental Settings}
\vspace{-5pt}\paragraph{Dataset} We experiment with \model{} on both class-conditioned and text-to-image generation tasks. For the former, we conduct experiments on ImageNet-1K~\citep{Deng2009ImageNet:Database} including $256 \times 256$ and $512 \times 512$ resolutions. For text-to-image, we show two settings: a constrained setting CC12M~\citep{changpinyo2021cc12m}, where each image is accompanied by a synthetic caption following~\citep{gu2024kaleido}. We also demonstrated a scaled setting where our models trained an in-house dataset with CC12M, in total $\sim 700$M text-image pairs.
%In addition, we also adopt the validation set of COCO2017 for testing zero-shot text-to-image generation.
\vspace{-5pt}\paragraph{Evaluation} In line with prior works, we report Fr\'echet Inception Distance (FID)~\citep{fid} %and Inception Score (IS)
to quantify the the realism and diversity of generated images. For text-to-image generation, we use MSCOCO 2017~\citep{Lin2014} validation set to assess the zero-shot capabilities of these models. We also report additional evaluation (e.g., GenEval~\citep{ghosh2023geneval}) in \cref{app:C}.
\vspace{-5pt}\paragraph{Model and Training Details}
We implement all models following the setup of~\citet{dubey2024llama}, using RoPE~\citep{su2024roformer} for positional encoding. By default, we set the architecture to $d(N) = 18(6)$ with a model dimension of 2048 (XL) and $24(6)$ with a dimension of 3096 (XXL) for class-conditioned and text-to-image models, respectively (\cref{sec.arch}), resulting in 1.4B and 3.8B parameters. Since \model{} operates in a compressed latent space, we are able to train all models with a patch size of $p=1$. For text-to-image models, we use T5-XL~\citep{raffel2020exploring} as the text encoder. To showcase the generality of our approach, we also train a variant where the deep block is initialized from a pretrained LLM (Gemma2~\citep{team2024gemma} in this case), without additional text encoder.
All models are pre-trained at $256 \times 256$ resolution on 400M images with a global batch size of 512. High-resolution finetuning is done by increasing input length. For text-to-image models, variable-length inputs are supported via mixed-resolution training: images are pre-classified into 9 shape buckets and flattened into sequences for unified processing. See \cref{app:B} for detailed settings.
\input{tables/imagenet_image}
\input{tables/text2image_image}
\input{tables/image_edit}
\input{tables/imagenet}
\input{tables/comparison_image}
\input{tables/ablation_image}
\subsection{Results}
\vspace{-5pt}\paragraph{Comparison with Baselines}
We benchmark our approach on class-conditioned ImageNet-256, comparing against diffusion and autoregressive models across both discrete and continuous domains (\cref{tb:imagenet}). For fair comparison, we train a TARFlow model~\cite{zhai2024normalizing} in pixel space with a similar parameter count and original architecture (8 flows, 8 layers each, width 1280). We also train a variant with our \emph{deep-shallow} design, identical to \model{} except for using pixel inputs with linearly scaled patch sizes. Among NF models, the \emph{deep-shallow} architecture consistently outperforms the standard design, and switching to latent-space inputs yields further gains.
Our method achieves competitive results compared to other baselines (\cref{tb:imagenet,tb:imagenet512}). Note the FID on ImageNet $256\times 256$ is near saturated to the upper-bound of the finetuned decoder (see additional details in \cref{app:B}).
Zero-shot evaluations on COCO (\cref{tab:model-comparison}) show strong performance on text-conditioned generation, demonstrating that NFs can also serve as a scalable and competitive generative modeling framework.
\vspace{-5pt}\paragraph{Qualitative Results}
\cref{fig:imagenet_sample} and \cref{fig:text2image_sample} present representative class- and text-conditioned generations, respectively. Our method delivers high-resolution images over a wide range of aspect ratios, with perceptual quality comparable to state-of-the-art diffusion and autoregressive approaches. \cref{fig:image_edit_sample} also highlights our model’s support for \textbf{image editing}. Further qualitative and interactive editing results appear in Appendix \ref{app:E}, underscoring the breadth and fidelity of our outputs.
\vspace{-5pt}
\paragraph{Comparison with Diffusion and Autoregressive Models}
We further compare \model{} with diffusion and autoregressive (AR) models to analyze training dynamics. \cref{fig:vs_ar_dit} shows FID trajectories using nearly identical architectures. While the FID gap between \model{} and the baselines is smaller when computed over 4,096 samples, \model{} consistently achieves the lowest FID at every training checkpoint when evaluated with 50,000 samples. This suggests that \model{} produces more diverse outputs, which may not be fully captured with smaller evaluation sets.
\cref{fig:vs_ar_dit} also compares inference throughput on a single H100 GPU for diffusion, AR, and \model{} models. Diffusion’s wall-clock time grows linearly with its number of refinement steps—$\approx 250$ steps at best FID—so it’s the slowest. By contrast, each step in AR and \model{} is only a lightweight forward pass whose per-token cost is low, allowing throughput to rise as batch size increases. Beyond a batch size of 32, \model{} outperforms the AR baseline by restricting guidance to the deep block and removing the per-token multinomial sampling loop, yielding superior inference-time scalability.
\vspace{-5pt}
\paragraph{Comparison of CFG Strategies}
As shown in the \cref{fig:vs_cfg}, the original strategy used in \citet{zhai2024normalizing} exhibits a sharp “dip‐and‐spike” behavior: it achieves its best FID at similar guidance weight as the newly proposed CFG, but then degrades quickly as you move away from that optimum.
Even when using the ``annealing trick''~\citep{zhai2024normalizing}, performance still suffers dramatically both scales. By contrast, our proposed CFG not only improves on the original’s best point—without additional tricks—but—more importantly—maintains nearly the same quality over a much wider range of guidance weights, which gives more flexibility in tuning text-conditioned generation tasks.
\vspace{-5pt}
\paragraph{Scalability Analysis} To assess the scalability, we perform a study by varying the depth of the deep block and tracking performance over training. \cref{fig:nll_model_size} reports negative log-likelihood (NLL) and \cref{fig:fid_model_size} shows FID with $4096$ samples across iterations. Both metrics indicate that deeper models converge faster and achieve better final performance, demonstrating the increased capacity.
\vspace{-5pt}
\paragraph{Ablation on Model Design}
To validate the theoretical insights from~\cref{prop:expressive}, we study how model expressivity varies with the number of layers $T$ in the deep block. Performance drops sharply when $T < 2$, while models with $T \geq 2$ perform similarly—consistent with~\cref{prop:expressive}. We also ablate the number and depth of deep blocks in \cref{fig:ablation_deep,fig:ablation_total}, finding that block depth is more critical than quantity, providing practical guidance for architectural design.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/imagenet_main_example1.jpeg}
\caption{\footnotesize Random samples of \model{} on ImageNet $256\times 256$ and $512\times 512$ ($\omega=3.0$).}
\label{fig:imagenet_sample}
\vspace{-5pt}
\end{figure}
%\FloatBarrier % makes sure earlier floats are placed
\begin{figure}[p]
\centering
\includegraphics[width=\linewidth]{figures/t2i_example_v3.pdf}
\caption{\footnotesize Selected samples of various aspect ratios from \model{} on for text-to-image generation ($\omega=4.0$). Image resolutions are adjusted proportionally for the ease of visualization.}
\label{fig:text2image_sample}
\vspace{-10pt}
\end{figure}
%\FloatBarrier % optional, to prevent floats jumping backward
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/image_editing_example.jpeg}
\caption{\footnotesize Example of Image editing using \model{}. Given an input image and simple description, our model can seamlessly edit the contents based on various instruction using with the learned model prior.}
\label{fig:image_edit_sample}
\vspace{-5pt}
\end{figure}
% \begin{table}[t]
% \caption{Performance of class‑conditional generative models on ImageNet ($256\times 256$).
% $^{*}$Models implemented only in pixel space. $^{**}$MAR relies on mask‑based iterative refinement rather than the KV‑cache typical of autoregressive transformers.
% }
% \label{sample-table}
% \begin{center}
% \begin{tabular}{l r r}
% \toprule
% Model & FID$\downarrow$ & \#params \\
% \midrule
% \textbf{Diffusion Models} \\
% ADM~\citep{dhariwal2021diffusion}$^{*}$ & 10.94 & 554M \\
% CDM~\citep{ho2022cascaded}$^{*}$ & 4.88 & - \\
% LDM~\citep{rombach2022high} & 3.60 & 400M \\
% RIN~\citep{jabri2022scalable}$^{*}$ & 3.76 & 410M \\
% DiT~\citep{peebles2023scalable} & 2.27 & 675M \\
% SiT~\citep{ma2024sit} & 2.06 & 675M \\
% %\jg{more baselines} \\
% \midrule
% \textbf{Autoregressive Models (discrete)} \\
% VQGAN~\citep{taming} & 15.78 & 1.4B \\
% RQTran~\citep{lee2022autoregressive} & 3.80 & 3.8B \\
% LlamaGen-XXL~\citep{sun2024autoregressive} & 2.34 & 775M \\
% LlamaGen-3B~\citep{sun2024autoregressive} & 2.18 & 3.1B \\
% VAR~\citep{tian2024visual} & 1.73 & 2.0B \\
% %ELM~\citep{liu2024elucidating} & 1.53 & 1.9B \\
% \midrule
% \textbf{Autoregressive Models (continuous) } \\
% Jetformer~\citep{tschannen2024jetformer}$^{*}$ & 6.64 & 2.75B \\
% MAR-AR~\citep{li2024autoregressive} & 4.69 & 479M \\
% MAR~\citep{li2024autoregressive}$^{**}$ & 1.55 & 943M \\
% DART~\citep{gu2024dart} & 3.82 & 820M \\
% GIVT~\citep{tschannen2024givt} & 2.59 & - \\
% \midrule
% \textbf{Normalizing Flow} \\
% TARFlow~\citep{zhai2024normalizing}$^{*}$ & 5.56 & 1.4B \\
% %\model{}(pixel) (Ours) \\
% \model{} (Ours) & 2.41 & 1.4B \\
% \bottomrule
% \end{tabular}
% \end{center}
% \label{tb:imagenet}
% \end{table}
% -------------------------------------------------------------------------
\begin{table*}[t]
\centering
% ===== LEFT column =======================================================
\begin{minipage}[t]{0.48\textwidth}
% --- (a) ImageNet table -------------------------------------------------
\setlength{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1.05}
\small
\captionof{table}{\footnotesize Class-cond ImageNet $256{\times}256$ (FID-50K)}
\vspace{-4pt}
\label{tb:imagenet}
\begin{tabular}{@{}lrr@{}}
\toprule
Model & FID$\downarrow$ & \# Param. \\
\midrule
\textbf{Diffusion Models} \\
ADM~\citep{dhariwal2021diffusion} & 10.94 & 554M \\
CDM~\citep{ho2022cascaded} & 4.88 & – \\
LDM~\citep{rombach2022high} & 3.60 & 400M \\
RIN~\citep{jabri2022scalable} & 3.76 & 410M \\
DiT~\citep{peebles2023scalable} & 2.27 & 675M \\
SiT~\citep{ma2024sit} & 2.06 & 675M \\
\midrule
\textbf{Autoreg. (discrete)} \\
VQGAN~\citep{taming} & 15.78 & 1.4B \\
RQTran~\citep{lee2022autoregressive} & 3.80 & 3.8B \\
%LlamaGen-XXL~\citep{sun2024autoregressive} & 2.34 & 775M \\
LlamaGen-3B~\citep{sun2024autoregressive} & 2.18 & 3.1B \\
VAR~\citep{tian2024visual} & 1.73 & 2.0B \\
\midrule
\textbf{Autoreg. (continuous)} \\
Jetformer~\citep{tschannen2024jetformer} & 6.64 & 2.75B \\
MAR-AR~\citep{li2024autoregressive} & 4.69 & 479M \\
MAR~\citep{li2024autoregressive} & 1.55 & 943M \\
DART~\citep{gu2024dart} & 3.82 & 820M \\
GIVT~\citep{tschannen2024givt} & 2.59 & – \\
\midrule
\textbf{Normalizing Flow} \\
TARFlow~\citep{zhai2024normalizing} \footnote{\small Implemented using their official codebase.} & 5.56 & 1.3B \\
TARFlow + \emph{deep-shallow} & 4.69 & 1.4B \\
\model{} (Ours) & 2.40 & 1.4B \\
\bottomrule
\end{tabular}
\end{minipage}%
\hfill
% ===== RIGHT column ======================================================
\begin{minipage}[t]{0.48\textwidth}
\setlength{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1.05}
\small
\captionof{table}{\footnotesize Class-cond ImageNet $512{\times}512$ (FID-50K)}
\vspace{-4pt}
\label{tb:imagenet512}
\centering
\begin{tabular}{@{}lcc@{}}
\toprule
Model & FID$\downarrow$ & \# Param. \\
\midrule
ADM-U~\citep{dhariwal2021diffusion} & 3.85 & 731M \\
DiT-XL/2~\citep{peebles2023scalable} & 3.04 & 674M \\
LEGO~\citep{zheng2024learning} & 3.74 & 681M \\
MaskDiT-G~\citep{Zheng2024MaskDiT} & 2.50 & 730M \\
EDM2-XXL\citep{karras2024guiding} &1.25 & 1.5B \\
\midrule
STARFlow (Ours) & 3.00 & 1.4B \\
\bottomrule
\end{tabular}
\vfill
\setlength{\tabcolsep}{1pt}
\renewcommand{\arraystretch}{1.05}
\small
\captionof{table}{\footnotesize Zero-shot T2I on COCO (FID-30K)}
\label{tab:model-comparison}
\begin{tabular}{@{}lrr@{}}
\toprule
Method & FID$\downarrow$ & \# Param. \\
\midrule
%\textbf{Autoregressive Models} & & \\
DALL·E~\citep{ramesh2021zero} & 27.5 & 12B \\
% LAFITE~\citep{zhou2022towards} & 26.9 & 75M \\
CogView2~\citep{ding2021cogview} & 24.0 & 6B \\
% Parti-750M~\citep{yu2022scaling} & 10.71 & 750M \\
% Parti-3B~\citep{yu2022scaling} & 8.10 & 3B \\
Make-A-Scene~\citep{makeascene} & 11.8 & – \\
DART~\citep{gu2024dart} & 11.1 & 800M \\
DALL·E 2~\citep{ramesh2022hierarchical} & 10.4 & 5.5B \\
GigaGAN~\citep{kang2023scaling} & 9.1 & 1B \\
%\textbf{Masked Models} & & \\
Muse~\citep{chang2023muse} & 7.9 & 3B \\
%\midrule
%\textbf{Diffusion Model} & & \\
% GLIDE~\citep{nichol2021glide} & 12.24 & 5B \\
%LDM-KL-8-G*~\citep{rombach2022high} & 12.63 & 1.45B\\
Imagen~\citep{ho2022imagen} & 7.3 & 3B \\
Parti-20B~\citep{yu2022scaling} & 7.2 & 20B \\
eDiff-I~\citep{balaji2022ediffi} & 7.0 & 9B \\
% UFOGen~\citep{Xu2023UFOGenYF} & 12.78 & 0.9B \\
% SiD-LSG~\citep{zhou2024long} & 8.15 & 0.9B \\
% DMD~\citep{Yin2023OnestepDW} & 11.49 & 0.9B \\
%\midrule
%\textbf{GANs} & & \\
% StyleGAN-T~\citep{sauer2023stylegan} & 13.90 & 1B \\
\midrule
%\textbf{NFs} & & \\
\model{}-CC12M & 10.3 & 3.8B \\
\model{}-CC12M-Gemma & 11.4 & 2.4B \\
\model{}-FullData & 9.1 & 3.8B\\
\bottomrule
\end{tabular}
\vspace{-0.8em} % space between the two right-hand tables
% --- (b) tiny placeholder table ----------------------------------------
\end{minipage}
\vspace{-14pt}
\end{table*}
% \begin{table}[t]
% \caption{Performance of class‑conditional generative models on ImageNet ($256\times 256$).
% $^{*}$Models implemented only in pixel space. $^{**}$MAR relies on mask‑based iterative refinement rather than the KV‑cache typical of autoregressive transformers.
% }
% \label{sample-table}
% \begin{center}
% \begin{tabular}{l l c c c c}
% \toprule
% Type & Model & FID$\downarrow$ & IS$\uparrow$ & \#params \\
% \midrule
% \multirow{5}{*}{Diffusion} & ADM~\citep{dhariwal2021diffusion}$^{*}$ & 10.94 & 101.0 & 554M \\
% & CDM~\citep{ho2022cascaded}$^{*}$ & 4.88 & 158.7 & - \\
% & LDM~\citep{rombach2022high} & 3.60 & 247.7 & 400M \\
% & RIN~\citep{jabri2022scalable}$^{*}$ & 3.76 & 182.0 & 410M \\
% & DiT~\citep{peebles2023scalable} & 2.27 & 278.2 & 675M \\
% & SiT~\citep{ma2024sit} & 2.06 & 277.5 & 675M \\
% %\jg{more baselines} \\
% \midrule
% \multirow{5}{*}{AR (Discrete)} & VQGAN~\citep{taming} & 15.78 & 74.3 & 1.4B \\
% & RQTran~\citep{lee2022autoregressive} & 3.80 & 323.7 & 3.8B \\
% & LlamaGen-XXL~\citep{sun2024autoregressive} & 2.34 & 253.9 & 775M \\
% & LlamaGen-3B~\citep{sun2024autoregressive} & 2.18 & 263.3 & 3.1B \\
% & VAR~\citep{tian2024visual} & 1.73 & 350.2 & 2.0B \\
% & ELM~\citep{liu2024elucidating} & 1.53 & 332.9 & 1.9B \\
% \midrule
% \multirow{5}{*}{AR (Continuous)}
% & Jetformer~\citep{tschannen2024jetformer}$^{*}$ & 6.64 & - & 2.75B \\
% & MAR-AR~\citep{li2024autoregressive} & 4.69 & 244.6 & 479M \\
% & MAR~\citep{li2024autoregressive}$^{**}$ & 1.55 & 303.7 & 943M \\
% & DART~\citep{gu2024dart} & 3.82 & 263.8 & 820M \\
% & GIVT~\citep{tschannen2024givt} & 2.59 & - & - \\
% % \multirow{3}{*}{Reprod.} & DiT*~\citep{peebles2022scalable} & 19.52 & 125.9 & 675M & 16 \\
% % % & DiT*~\citep{peebles2022scalable} (cfg=2.0) & 7.75 & 225.16 & 675M & 16 \\
% % & SiT*~\citep{ma2024sit} & 6.98 & 122.9 & 675M & 16 \\
% % % & SiT*~\citep{ma2024sit} (cfg=2.0) & 3.80 & 341.72 & 675M & 16 \\
% % & MAR*~\citep{li2024autoregressive} & 6.37 & 221.3 & 943M & 16 \\
% \midrule
% \multirow{2}{*}{NF} & TARFlow~\citep{zhai2024normalizing}$^{*}$ & 5.56 \\
% & \model{}(pixel) (Ours) \\
% & \model{} (Ours) & $\sim$2.4 & & 1.4B \\
% \bottomrule
% \end{tabular}
% \end{center}
% \label{tb:imagenet}
% \end{table}
\begin{figure}[t]
\centering
\begin{subfigure}[b]{\textwidth}
\includegraphics[width=0.32\linewidth]{figures/fid_plot.png}\hfill
\includegraphics[width=0.32\linewidth]{figures/fid_plot50k.png}\hfill
\includegraphics[width=0.32\linewidth]{figures/speed_plot.png}
\caption{\label{fig:vs_ar_dit}\footnotesize Generation Quality (4096 and 50K samples) and Speed Comparison with DiT and AR}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\linewidth]{figures/cfg_fid_plot.png}
\caption{\label{fig:vs_cfg}\footnotesize Original vs.\ Proposed CFG}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\linewidth]{figures/size_nll_plot.png}
\caption{\label{fig:nll_model_size}\footnotesize NLL vs.\ Model size}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\linewidth]{figures/size_fid_plot.png}
\caption{\label{fig:fid_model_size}\footnotesize FID vs.\ Model size}
\end{subfigure}
\vspace{0.5em}
\begin{subfigure}[b]{0.65\textwidth}
\includegraphics[width=\linewidth]{figures/model+config1.png}
\caption{\label{fig:ablation_deep}\footnotesize Impact of varying the number of layers within deep blocks.}
\end{subfigure}\hfill%
\begin{subfigure}[b]{0.32\textwidth}
\includegraphics[width=\linewidth]{figures/model_config2.png}
\caption{\label{fig:ablation_total}\footnotesize Ablation of parameter allocation.}
\end{subfigure}
\caption{\footnotesize Experimental results of comprehensive ablation study}
\label{fig:all_metrics}
\vspace{-10pt}
\end{figure}
% \begin{figure}[t]
% \centering
% \begin{subfigure}[b]{0.65\textwidth}\includegraphics[width=\linewidth]{figures/model+config1.png}\caption{Ablation1}\end{subfigure}\hfill%
% \begin{subfigure}[b]{0.32\textwidth}\includegraphics[width=\linewidth]{figures/model_config2.png}\caption{Ablation2}\end{subfigure}
% \caption{Comparison of model metrics.}
% \label{fig:ablation}
% \end{figure}
\vspace{-0.2cm}
\section{Related Work}
\vspace{-5pt}\paragraph{Continuous Normalizing Flows, Flow Matching, and Diffusion Models}
Normalizing Flows (NFs) can be extended to continuous-time via Continuous Normalizing Flows (CNFs)~\citep{chen2018neural}, which model transformations as ODEs. This relaxes the need for explicit invertible mappings and simplifies Jacobian computation to a trace~\citep{grathwohl2018ffjord}, though it requires noisy stochastic estimators~\citep{hutchinson1989stochastic}. Flow Matching~\citep{lipman2023flow}, inspired by CNFs, learns sample-wise interpolations between prior and data using vector fields grounded in Tweedie’s Lemma~\citep{efron2011tweedie}. While CNFs and NFs optimize exact likelihoods through invertible mappings, Flow Matching aligns more closely with diffusion models, sharing variational training objectives.
\vspace{-5pt}\paragraph{Autoregressive Models}
Discrete autoregressive models, especially large language models~\citep{gpt3,dubey2024llama,guo2025deepseek}, dominate modern generative AI by scaling next-token prediction. Scaling laws~\citep{kaplan2020scaling} show predictable gains with more data and parameters. These models now power leading multimodal systems for both understanding and generation~\citep{liang2024survey,sun2024autoregressive,tian2024visual,li2025fractal}.
To overcome information loss from quantization, recent work extends AR modeling to continuous spaces, using mixture-of-Gaussians~\citep{tschannen2024givt,tschannen2024jetformer} or diffusion decoding~\citep{li2024autoregressive,gu2024dart,fan2024fluid}. Hybrid approaches also emerge, unifying AR and diffusion paradigms~\citep{gu2024kaleido,zhou2024transfusion,openai2024gpt4o}.
\section{Conclusion and Limitation}
We have presented \model{}, the first latent based normalizing flow model that scales to high resolution images and large scale text to image modeling. Our results demonstrate that normalizing flows are scalable generative modeling method, and is capable of achieving comparable results to strong diffusion and autoregressive baselines.
There are also limitations to our work. For example, we have exclusively relied on pretrained autoencoders for simplicity, but it leaves the question of a potential joint latent–NF model design unexplored. Moreover, in this work we have primarily focused on training high-quality models, which comes at the cost of un-optimized inference speed. Additionally, our evaluation has been restricted to class- and text-conditional image generation on standard benchmarks; how well the approach generalizes to other modalities (e.g., video, 3D scenes) or more diverse, real-world data distributions remains to be seen.
\section*{Acknowledgements}
We thank Ying Shen, Yizhe Zhang, Navdeep Jaitly, Alaa El-Nouby and Preetum Nakkiran for helpful discussions.
We also thank Samy Bengio for leadership support that
made this work possible.
% 1. Big, centred “Appendices” title page
\newpage
\appendix
\part*{\huge Appendix}
\section{Derivations}\label{app:A}
\subsection{Extended Discussion of \cref{prop:expressive}}
\input{tables/prop1}
\subsection{Proof of \cref{prop:new_cfg}}
\input{tables/prop2}
\section{Implementation Details}\label{app:B}
\subsection{Architecture Design}
\label{app:config}
\vspace{-5pt}\paragraph{Overall Structure.}
We implement \model{} with a decoder-only Transformer~\citep{transformer}.
The shorthand $l(N)\!-\!d$ (see \cref{sec.arch}) denotes a single \emph{deep}
AF block of $l$ layers followed by $N{-}1$ \emph{shallow} blocks (two layers
each) with hidden width $d$. Our class-conditioned baseline uses
$18(6)\!-\!2048$ ($\approx$1.4 B parameters), while the text-conditioned model
uses $24(6)\!-\!3072$ ($\approx$3.8 B parameters). Layer-allocation sweeps in
\cref{fig:all_metrics}(b–e) probe scalability and convergence. Unlike \citet{zhai2024normalizing}, we apply a final layer norm at the predictions of each Transformer block.
\vspace{-5pt}\paragraph{Conditioning Mechanism.}
For both conditioning modes, the context is prepended as a prefix to the deep
block, and we omit AdaLN~\citep{peebles2023scalable}—a choice that simplifies
the network and marginally improves quality. ImageNet classes are provided as
one-hot vectors. Text captions (T2I) are encoded by a frozen
\textsc{Flan-T5-XL} encoder~\citep{raffel2020exploring}, truncated to 128
tokens.
\vspace{-5pt}\paragraph{VAE Latent Space.}
Images are first mapped to continuous latent tokens via the DiT
VAE~\citep{peebles2023scalable}, which compresses spatial dimensions by
$48\times$. Because performance is highly sensitive to patch size, we keep
$p=1$ for all resolutions, yielding sequences of
$1024$, $4096$, and $16384$ tokens for
$256\times256$, $512\times512$, and $1024\times1024$ images, respectively.
We also applied our proposed deep-shallow architecture in pixels (see \cref{tab:model-comparison}). To match similar computation, we adopted a patch size of $p=8$ for learning $256\times 256$ images.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/app_cfg_exp.jpeg}
\caption{\footnotesize (a) Direct generation results using the model’s latent samples without decoder fine-tuning or score-based denoising. (b) Results after applying decoder fine-tuning, effectively reducing latent-space noise. (c) and (d) provide comparisons of classifier-free guidance (CFG) strategies for text-to-image generation: (c) demonstrates degraded outputs at guidance weight $\omega=5$ using the approach of \citet{zhai2024normalizing}, whereas (d) shows stable results with our proposed CFG method, confirming its improved effectiveness and suitability for text-conditioned applications.}
\label{fig:app_cfg_exp}
\end{figure}
\vspace{-5pt}\paragraph{Positional Embeddings.}
All variants employ rotary positional embeddings (RoPE)~\citep{su2024roformer};
we adopt \textbf{3D-RoPE}, giving each token $(x,y,t)$, where $(x,y)$ encodes
its spatial grid location ($(0,0)$ for text tokens) and $t$ its caption index
($0$ for image tokens). During fine-tuning from $256\times256$ to higher
resolutions, we align positions by setting
$(x',y',t) \!=\! (x/\alpha,\,y/\alpha,\,t)$, where $\alpha$ is the
up-sampling ratio.
\vspace{-5pt}\paragraph{Default Configuration.}
Below is the default configurations of \model{}s:
\begin{verbatim}
model config for \model{}-l(N)-d:
patch_size=1
hidden_size=d
num_layers=[l] + [2] * (N-1)
num_channels_per_head=64
use_swiglu_ffn=False
use_rope=True
use_final_rmsnorm=True
\end{verbatim}
\subsection{Training Details}
In all the experiments, we share the following training configuration for our proposed \model{}. Models are trained on $32$ (for 1.4B model) or $64$ (for 3.8B model) H100 GPUs for around $2$ weeks.
\begin{verbatim}
training config:
batch_size=512
optimizer='AdamW'
adam_beta1=0.9
adam_beta2=0.95
adam_eps=1e-8
learning_rate=1e-4
min_learning_rate=1e-6
learning_rate_schedule=cosine
weight_decay=1e-4
max_training_images=400M
mixed_precision_training=bf16
\end{verbatim}
\vspace{-5pt}\paragraph{Stability of \cref{eq:tarflow_loss}.}
The maximization term $-\log\sigma$ in \cref{eq:tarflow_loss} is
\emph{unbounded}: the model can drive some $\sigma$ values arbitrarily close to
zero whenever this hardly influences $\vz$, echoing a classic pathology of
normalizing-flow training. We mitigate it with three safeguards:
\begin{enumerate}[leftmargin=*]
\item \textbf{Soft clipping.} Each raw Transformer output $\mathbf{x}$ is
mapped through
\(f(\mathbf{x}) = a\,\tanh\bigl(\mathbf{x}/a\bigr)\), softly limiting
its magnitude to $\pm a$.
\item \textbf{Positive scale parameterization.} The scale is
enforced positive via
\(\sigma = \operatorname{softplus}\bigl(\hat{\sigma}\bigr)\), where
$\hat{\sigma}$ is the network’s variance output.
\item \textbf{Latent norm penalty.} We add a small norm penalty over the intermediate latents $\vx^t$ to avoid extremely large values. Typically a weight of 1e-4 is enough to keep the magnitude stable without hurting the performance.
\end{enumerate}
\input{tables/geneval}
\paragraph{Mixed-Resolution Training.}
During the high-resolution phase, \model{} supports \emph{mixed resolutions},
preserving each image’s native aspect ratio. Because the backbone is a
Transformer, variable sequence lengths are handled naturally, so no aggressive
cropping is required; this better retains scene content and improves
caption–image alignment. We bucket images into nine aspect-ratio bins:
$21{:}9$, $16{:}9$, $3{:}2$, $5{:}4$, $1{:}1$, $4{:}5$, $2{:}3$, $9{:}16$,
and $9{:}21$ with the ratio appened in the caption:
\begin{verbatim}
{original_caption}\n in a {aspect_ratio} aspect ratio.
\end{verbatim}
Image is center-cropped and resized so that its token
count roughly matches that of a square reference. For a
$512\times512$ target, we enforce $H\times W\!\approx\!512^{2}$. This
procedure stabilizes optimization, maximizes GPU utilization, and is used in
conjunction with the 3D-RoPE alignment described above.
\subsection{Decoder Fintuning Details}
We perform decoder fine-tuning by freezing the encoder and introducing controlled noise into the latent representations. The decoder is then trained using a standard autoencoder loss comprising L2, perceptual, and GAN losses. Training is conducted on ImageNet images at a resolution of 256×256 for 200K updates with a batch size of 64, utilizing a single node with 8 GPUs. To monitor performance, we compute an rFID by randomly sampling 50K real images, adding Gaussian noise with a standard deviation of 0.3, and directly decoding these perturbed images to compare with real images. Our resulting rFID is approximately 2.73, which exceeds the best achievable gFID from \model{} at 2.40. This suggests current \model{} with this finetuned decoder might have reached a performance ceiling under the specified noise conditions, highlighting an avenue for future exploration. Notably, although trained only on ImageNet at $256\times 256$ resolution, the fine-tuned decoder can seamlessly generalize to arbitrary resolutions, aspect ratios, and text-to-image domains.
See visual comparison in \cref{fig:app_cfg_exp} (a) and (b).
\subsection{Baseline Details}
\paragraph{Diffusion Model Baseline}
We deploy the official implementation of DiT\footnote{\url{https://github.com/facebookresearch/DiT}} and report the performance.
To make the architecture comparable to \model{}, we set the number of layers to 28 and hidden dimension to 2048 while keeping the number of attention heads to 16, resulting in a model size of 2.1B parameters.
We kept all of the other official repository settings the same.
Notably the pretrained VAE of the official repository matches the one used in \model{}.
The baseline DiT is trained for 200M samples with batch size 256 using the official implementation settings: AdamW optimizer with learning rate $0.0001$ and no weight decay $0.0$.
In inference, we set the number of sampling steps to $250$ and classifier-free guidance scale to $1.5$ following the best reported setting in the original paper.
\paragraph{Autoregressive Model Baseline}
We deploy the official implementation of LlamaGen\footnote{\url{https://github.com/FoundationVision/LlamaGen}}~\citep{sun2024autoregressive} and report the performance. In particular, to make the architecture comparable to our \model{}, we set the number of layers as 28, hidden dimension 2048, and number of attention heads 32, which leads to the total model size of 1.4B parameters. We also adopt the VQ-VAE from the official repository with downsample factor 8 which matches the downsample factor used in \model{}. The baseline LlamaGen is trained for 200M samples with batch size 512 using AdamW optimizer with learning rate $0.0001$, weight decay $0.05$ and betas $(0.9, 0.95)$. In inference, we set the top-k the same as the vocabulary size $16384$ and temperature $1.0$. We also implement classifier-free guidance with scale $1.75$ following the best reported setting in the original paper.
\section{Additional Experiments}\label{app:C}
\subsection{Additional Evaluation on Text-to-Image Generation}
\cref{tab:method_comparison} summarizes our GenEval~\citep{ghosh2023geneval} performance against representative diffusion and autoregressive (AR) baselines. \model{} attains an Overall score of \textbf{0.56}—slightly above SDXL (0.55) and well ahead of earlier Stable Diffusion checkpoints—while simultaneously surpassing the several recent AR models for text-to-image generation, including Emu-3 (0.54), Chameleon (0.39), and LlamaGen (0.32). Improvements are most pronounced on the more compositional sub-tests.
Crucially, these gains are achieved \emph{without} any reward-based alignment, target-dataset finetuning, or caption rewriting—\model{} is trained once, end-to-end, and evaluated exactly as generated.
Because GenEval isolates visual grounding, we purposefully restrict comparison to image-only generators; nonetheless, \model{}’s single-pass inference already delivers substantial latency advantages over diffusion models that require tens to hundreds of denoising steps. The availability of exact log-likelihoods further opens avenues for principled preference learning, sequential planning, or cascaded generation—capabilities that likelihood-free baselines lack. An exciting next step is to couple \model{} with large pretrained language- or vision–language models, forming a unified system that reasons jointly over text and images while retaining the speed, stability, and strong grounding demonstrated here.
\subsection{Inference Speed}
Because \model{} is autoregressive, tokens must be generated sequentially through every AF block, which makes inference latency the dominant bottleneck.
Our deep–shallow redesign \textbf{partially} mitigates this issue: by concentrating parameters in the first few “deep’’ blocks and leaving the remaining ones lightweight with no condition or guidance, the incremental cost of later blocks becomes minimal.
In practice, while the sampling speed is still relatively slow, this layout also outperforms the equal-sized architecture of \citet{zhai2024normalizing} (see \cref{tab:inference_speed}), and its overall runtime approaches that of a standard LLM—leaving additional head-room for techniques such as distillation or speculative decoding.
\begin{table}[t]
\centering
\caption{\footnotesize Per-block inference time (s) with a fixed batch size $16$ of the 1.4B sized model for generating $256\times 256$ images. Sampling speed is measured with CFG.
The proposed deep–shallow uses \textbf{6} blocks: an 18-layer Transformer followed by a 5 blocks of 2-layer Transformer. The hidden dimension is $2048$.
The equal-sized~\citep{zhai2024normalizing} baseline uses \textbf{8} blocks where each block has 8 layer of Transformers. To match the overall parameters, we reduce the hidden dimensions to $1280$.
}
\begin{tabular}{lccccccccc}
\toprule
Block ID & 0 & 1 & 2 & 3 & 4 & 5 & 6 & 7 & Total (s) \\
\midrule
Equal-sized~\citep{zhai2024normalizing} & 9 & 9 & 9 & 9 & 9 & 9 & 9 & 9 & 72 \\
Deep–shallow (ours) & 18 & 2 & 2 & 2 & 2 & 2 & - & - & 35 \\
\bottomrule
\end{tabular}
\label{tab:inference_speed}
\end{table}
\subsection{Latent Denoising}
A second limitation is that \model{} cannot be trained directly on clean latents; adding Gaussian noise is required to keep the flow learning stable~\citep{ho2019flow++,zhai2024normalizing}, but this both complicates optimization and necessitates an explicit denoising stage at inference time.
In this work, we investigated three strategies of denoising:
\begin{enumerate}[nosep,leftmargin=*,label=(\arabic*)]
\item \textbf{Single-step score denoising}.
Use the flow itself as a score estimator and apply one denoising step.
Works only for mild noise; at $\sigma=0.3$ outputs are noticeably blurry.
\item \textbf{Multi-step diffusion denoising}.
Start from the noisy latent and run unconditional DDIM steps with a pretrained DiT.
Quality improves, but latency and model complexity increase substantially.
\item \textbf{Decoder finetuning (ours)}.
Finetune the VAE decoder so it can reconstruct directly from noisy latents. Training can be done very efficiently on unconditional images, and the GAN objective effectively handles the uncertainty.
This option is the simplest to deploy.
\end{enumerate}
\begin{table}[htbp]
\centering
\caption{\footnotesize Comparison of latent-denoising strategies at $\sigma_L=0.3$ on ImageNet $256\times 256$.}
\begin{tabular}{lcccc}
\toprule
Method & Extra Model & Extra Steps & FID 50K $\downarrow$ & Remarks \\
\midrule
Single-step score & -- & 1 & 2.96 & Blurry \\
Multi-step DiT (from 0.3) & DiT-XL & 30 & 2.53 & Slowest \\
Decoder finetune & Finetuned Decoder & 0 & \textbf{2.40} & Best, simplest \\
\bottomrule
\end{tabular}
\label{tab:latent_denoising}
\end{table}
Future work will aim for a principled solution that trains directly on clean data, eliminating the denoising stage entirely.
\section{Application Details}\label{app:D}
\subsection{Training-free Image Inpainting with \model{}}
Let $M\!\in\!\{0,1\}^{H\times W}$ be a binary mask that selects the pixels to
be filled and let $\vx_{\mathrm{gt}}\!\in\!\mathbb{R}^{H\times W\times C}$ be
the ground-truth image (available only at evaluation time for measuring
fidelity). We split the image into the \emph{observed} part
$\vx_{O}=(1-M)\odot\vx_{\mathrm{gt}}$ and the \emph{missing} part
$\vx_{M}=M\odot\vx_{\mathrm{gt}}$. The pretrained flow
$\vf_{\theta}:\vx\mapsto\vz$ induces a tractable density
$p_{\theta}(\vx)\!=\! \mathcal{N}(\vf_{\theta}(\vx);\,\mathbf{0},I)
\left|\det\nabla_{\!\vx}\vf_{\theta}\right|$. To sample from the conditional
$p_{\theta}(\vx_{M}\mid\vx_{O})$ \emph{without retraining}, we construct a
Metropolis–Hastings (MH) chain in latent space:
\begin{enumerate}[leftmargin=*]
\item \textbf{Init.}
Replace the missing region by Gaussian noise,
$\tilde{\vx}^{(0)} = \vx_O + M\odot\vepsilon,\;
\vepsilon\!\sim\!\mathcal{N}(0,\sigma^{2}I)$,
and map to latent space $\vz^{(0)}=\vf_{\theta}(\tilde{\vx}^{(0)})$.
\item \textbf{Proposal.}
Draw fresh noise in the \emph{same} masked region of latent space
\[
\vz' = \vz^{(t)} + M\odot\vgamma,\qquad
\vgamma\sim\mathcal{N}(0,\tau^{2}I),
\]
and obtain the candidate image $\vx'=\vf_{\theta}^{-1}(\vz')$.
We then restore the context pixels,
$\tilde{\vx}' = \vx_O + M\odot\vx'$,
ensuring every proposal satisfies the observed evidence.
\item \textbf{Acceptance.}
Because the forward and reverse proposals are symmetric, the MH
acceptance probability reduces to the ratio of conditional
probabilities:
\[
\alpha = \min\!\Bigl\{1,\;
\frac{p_{\theta}(\tilde{\vx}' \mid \vx_O)}
{p_{\theta}(\tilde{\vx}^{(t)} \mid \vx_O)}\Bigr\}
~=~ \min\!\Bigl\{1,\;
\exp\!\bigl[\log p_{\theta}(\tilde{\vx}')-
\log p_{\theta}(\tilde{\vx}^{(t)})\bigr]\Bigr\}.
\]
Accept with probability $\alpha$; otherwise keep the current state.
\item \textbf{Iteration.}
Set $t\!\leftarrow\!t+1$ and repeat steps (ii)–(iii) until convergence;
the final sample $\tilde{\vx}^{(T)}$ is reported as the inpainted image.
\end{enumerate}
Intuitively, each step perturbs only the masked latents, letting the powerful
flow prior propose content that is \emph{globally} coherent with the context
while the MH test enforces exact consistency with the joint density. The
chain is ergodic—Gaussian noise gives non-zero probability to every latent
configuration—and its stationary distribution is precisely
$p_{\theta}(\vx_{M}\!\mid\!\vx_{O})$.
In practice we set both $\sigma=1$ and $\tau=1$. Since our \model{} is well-trained on large-scale text-to-image data with sufficient capacity,
it yields high acceptance rates and we set the total iterations to $\bm{20}$. No additional training, guidance network, or
data-specific tuning is required.
effective plug-in for image inpainting with pretrained autoregressive flows.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/inpainting.pdf}
\caption{Demonstration of generation trajectories of inpainting output.}
\label{fig:app_editing}
\vspace{-20pt}
\end{figure}
\subsection{Interactive Image Editing with \model{}}
\model{} can be naturally extended to multi-round tasks such as interactive
image editing. We start from a pretrained text-to-image checkpoint and
finetune on the \textsc{AnyEdit} corpus%
\footnote{\url{https://dcd-anyedit.github.io}}.
For simplicity, we use only the subset that provides text instructions.
Each training quadruple
$(\vx^{\mathrm{src}},\,\vt^{\mathrm{cap}},\,\vt^{\mathrm{inst}},\,
\vx^{\mathrm{tgt}})$
contains a source image, its caption, a free-form instruction, and the edited
target (see \cref{fig:application_example}b).
We serialize every sample into the sequence
\[
\bigl[
\mathrm{T5}(\vt^{\mathrm{cap}}),\;
\mathrm{AFs}\!\bigl[\mathrm{VAE}(\vx^{\mathrm{src}})\bigr],\;
\mathrm{T5}(\vt^{\mathrm{inst}}),\;
\mathrm{AFs}\!\bigl[\mathrm{VAE}(\vx^{\mathrm{tgt}})\bigr]
\bigr],
\]
where image segments are tokenized by our VAE ($p=1$) and text segments are
embedded by a frozen \textsc{Flan-T5-XL}~\citep{raffel2020exploring,t5xl}.
Image latents first pass through the shallow-AF blocks independently, after
which all tokens are processed by the shared deep-AF Transformer. Because the
deep block is strictly causal, the edited image and all later tokens can
attend to the entire prefix—including the source image—without any special
masking. During inference the prefix is written once into the
\textit{KV} cache; sampling the edited tokens simply reads from this cache,
mirroring the behavior of language-only LLMs.
\vspace{-5pt}\paragraph{Joint Training Objective.}
Instead of optimizing a single conditional likelihood, we maximize the
\emph{joint} log-likelihood of both images:
\[
\max_{\theta}\,\mathcal{L}_{\mathrm{joint}}
~=~
\mathbb{E}_{(\vx^{\mathrm{src}},\vt^{\mathrm{cap}},\vt^{\mathrm{inst}},
\vx^{\mathrm{tgt}})}
\Bigl[
\log p_{\theta}\bigl(\vx^{\mathrm{src}} \mid \vt^{\mathrm{cap}}\bigr)
+
\log p_{\theta}\bigl(\vx^{\mathrm{tgt}}
\mid \vt^{\mathrm{inst}},\vx^{\mathrm{src}},\vt^{\mathrm{cap}}\bigr)
\Bigr],
\]
where each term is evaluated via the change-of-variables formula
(\cref{eq:tarflow_loss}). This objective maintains maximum-likelihood
training, allows gradients to propagate across \emph{all} modalities, and
enables the same network to generate from scratch (empty image prefix) or
perform edits (given image prefix).
Unlike diffusion-based MLLMs that first \emph{generate} pixels and then
re-encode them with a separate vision backbone, our autoregressive flow is
invertible: a single forward pass encodes the user image, and a single reverse pass decodes the edited result. Encoding and decoding share parameters,
introduce no information loss, and integrate seamlessly with the Transformer's \textit{KV} cache. This \emph{single-pass round-trip} property sharply
reduces latency and highlights autoregressive flows as a compelling choice for tightly coupled vision–language applications.
We show interactive image generation and editing examples in \cref{fig:app_editing} where given a caption and editing instruction, our model predicts two images one after another.
\section{Related Topic Discussion}
\subsection{Autoregressive Model v.s. Autoregressive Flow}
Connections between the two families emerge in masked autoregressive flows (MAF, \citep{papamakarios2017masked}), which impose invertibility on an autoregressive factorisation, yet fundamental differences remain. Autoregressive models dispense with any latent prior; each conditional distribution is learned directly in the data domain, which is typically discrete—tokens, integer pixels, or quantized audio samples—and generation proceeds strictly one element at a time. Normalizing flows, by contrast, begin from an explicit Gaussian prior in a continuous latent space and learn an invertible transformation that warps this prior into the target distribution. This design delivers exact log-likelihoods, parallel one-shot sampling, and bidirectional latent inference, but at the cost of enforcing invertibility and differentiability in every layer. While MAF narrows the gap by marrying an autoregressive factorisation with invertibility, the reliance on a Gaussian base and a continuous formulation remains the defining hallmark of normalizing flows, whereas the absence of a prior and the natural alignment with discrete data continue to characterise pure autoregressive models.
\subsection{Flow Matching v.s. Autoregressive Flow}
Normalizing flows (NF) and Flow Matching (FM) both map a simple latent prior to the data distribution, but they differ fundamentally in what they optimise and how they realise the map. A normalizing flow learns a time‐independent bijection whose parameters are updated by directly minimising the data’s negative log-likelihood (NLL); the change-of-variables formula provides an exact, unbiased gradient, so every parameter update moves the model toward the true maximum-likelihood solution. Flow Matching instead specifies a time‐dependent vector field that transports probability mass along a chosen path and trains this field with a velocity-matching loss. In short, Flow Matching reduces per-iteration cost by relaxing the objective, but Normalizing Flows retain the rigorous maximum-likelihood foundation, and exact densities.
% \subsection{Autoregressive Diffusion v.s. Autoregressive Flow}
\section{Broader Impacts}
\paragraph{Positive societal impacts.}
\model{} shows—for the first time—that \emph{normalizing flow}–based models can scale to the same resolutions and sample quality previously dominated by diffusion and discrete autoregressive methods.
The invertibility of normalizing flows enables interactive image editing (See Fig.~9b for examples) making \model{} suitable for assistive technologies (e.g., real-time diagram manipulation for education or accessibility) and for professional design workflows that demand faithful, iterative refinement.
\paragraph{Potential risks and negative impacts.}
Higher-quality image generation lowers the barrier to fabricating realistic—but false—visual evidence. Interactive editing magnifies this risk by enabling rapid revision cycles. We advocate the concurrent development of reliable flow-specific watermarking and provenance tools.
\section{Additional Samples}\label{app:E}
We show more generated samples from \model{} in \cref{fig:app_imagenet,fig:app_t2i,fig:app_t2i1024,fig:app_t2i1024v2}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/interactive_editing_1.1.jpeg}
\caption{\footnotesize Interactive editing with \model{}. Starting from an initial caption, \model{} generates a base image. Given a subsequent user-provided editing instruction, the model then modifies the image accordingly—without requiring re-encoding. Each example illustrates a generic instruction applied to a generated image. All images are synthesized at a resolution of $512\times512$. }
\label{fig:app_editing}
\vspace{-20pt}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/imagenet_samples_1.jpeg}
\caption{\footnotesize Additional class-conditioned generation from \model{}s trained on $256\times 256$ and $512\times 512$, respectively. The classes are \emph{sulphur-crested cockatoo, Kakatoe galerita, Cacatua galerita}, \emph{loggerhead, loggerhead
turtle, Caretta caretta}, and \emph{Siberian husky}.}
\label{fig:app_imagenet}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/t2i_samples_1.jpeg}
\caption{\footnotesize Additional text-conditioned generation from \model{}s trained on $256\times 256$ and $512\times 512$, respectively.}
\label{fig:app_t2i}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/t2i_example_1024.jpeg}
\caption{\footnotesize Additional text-conditioned generation from \model{}s trained on $1024 \times 1024$.}
\label{fig:app_t2i1024}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.97\linewidth]{figures/t2i_examples_variable.jpeg}
\caption{\footnotesize Additional text-conditioned samples from \model{}s trained on various aspect ratios.}
\label{fig:app_t2i1024v2}
\end{figure}
\begin{comment}
\begin{remarkbox}
\textbf{Prompts for \cref{fig:teaser}}
1. vw bus, canvas art, abstract art printing, in the style of brian mashburn, light red and light brown, theo prins, charming character illustrations, pierre pellegrini, vintage cut-and-paste, rusty debris --ar 73:92 --stylize 750 --v 6"
2. A red fox curled up asleep in a snowy woodland clearing, with delicate snowflakes falling gently around it, watercolor style.
3. A landscape featuring mountains, a valley, sunset light, wildlife and a gorilla, reminiscent of Bob Ross's artwork
4. Vibrant abstract painting with colorful swirling brushstrokes in blue, orange, red, yellow, and green, resembling a dynamic river of paint. In the center, bold hand-painted text reads “Go with the \model{}” in a playful and modern font. The style is fluid, energetic, and uplifting, like a digital acrylic artwork.
5. A rustic kitchen table set with freshly baked bread, an assortment of cheeses, a bowl of ripe fruit, and a bouquet of lavender, sunlight streaming through a nearby window casting soft shadows, detailed still life painting with warm tones and textures.
6. A panda eating bamboo in a lush green forest, with soft sunlight filtering through the leaves, realistic painting.
\end{remarkbox}
\end{comment}
\paragraph{Why a Single Block ($T=1$) Cannot Be Universal.}
With only one autoregressive–flow block,
\[
\vx_{d} \;=\; \mu_{\theta}(\vx_{<d}) \;+\; \sigma_{\theta}(\vx_{<d})\,
\vz_{d},
\qquad
\vz_{d}\sim\mathcal{N}(0,1),\; d=1,\dots,D,
\]
each conditional $p(\vx_{d}\!\mid\!\vx_{<d})$ is \emph{necessarily a single
Gaussian}. Because no latent variable influences the affine parameters
\emph{beyond} the current coordinate, the model cannot represent multimodal
densities or heavy tails. Consequently, $T=1$ flows are \emph{not} dense in
$L^{1}(\mathbb{R}^{D})$ and fail the universal-approximation criterion.
\paragraph{Why $T=2$ Is Almost Sufficient.}
For $T=2$ blocks with \emph{opposite} orderings, all coordinates except the
last ($d<D$) are expressed as infinite Gaussian mixtures
(\cref{eq.af_cond}) and hence enjoy the universal-approximation property via
the density of Gaussian mixtures~\citep{Goodfellow-et-al-2016}. The principal reason why $\vx_D$ fails to possess the universal approximation property lies in the structure:
\begin{align}
\vx_D = \mu_{\theta}^{b}(\vx_{<D}) + \sigma_{\theta}^{b}(\vx_{<D}) \cdot \vy_D,
\end{align}
where $\vy_D$ is defined as
\begin{align}
\vy_D &= \mu_{\theta}^{a}(\vy_{>D}) + \sigma_{\theta}^{a}(\vy_{>D}) \cdot \vz_D \\
&= \mu_{\theta}^{a}(\varnothing) + \sigma_{\theta}^{a}(\varnothing) \cdot \vz_D.
\end{align}
It is evident that $\vy_D$ follows a unimodal Gaussian, since $\vz_D$ is sampled from a unimodal Gaussian prior and the functions $\mu_{\theta}^{a}$ and $\sigma_{\theta}^{a}$ receive no random variable input, regardless of their nonlinearity. Consequently, $\vx_D$ also becomes a unimodal Gaussian, inheriting this limitation from $\vy_D$.
\paragraph{Why $T\!\ge 3$ Restores Universality.}
Introducing a \emph{third} block injects fresh latent variables that feed into
the affine parameters of the final coordinate. In effect, $\vx_{D}$ now
depends on a random input produced by the second block, exactly as $\vx_{d}$
($d<D$) depends on $\vy_{>d}$ in the $T=2$ proof. Consequently every
conditional $p(\vx_{d}\!\mid\!\vx_{<d})$ becomes an (infinite) Gaussian
mixture, and the entire joint density is dense in
$L^{1}(\mathbb{R}^{D})$. Additional blocks ($T>3$) only enlarge the model
class and do not degrade this property.
\smallskip
In summary, $T=1$ flows are fundamentally limited to unimodal Gaussians;
$T=2$ flows with alternating orderings achieve universality on
$D-1$ coordinates but leave the final one unimodal; and $T\!\ge3$ flows
overcome this last obstacle, granting full universal approximation power.
\begin{proof}
For an isotropic Gaussian $p(\vx)=\mathcal{N}(\mu,\sigma^{2}I)$ the score is
\[
\nabla_{\vx}\log p(\vx)\;=\;-\frac{\vx-\mu}{\sigma^{2}}.
\]
Hence
\[
\nabla_{\vx}\log p_c(\vx) = -\frac{\vx-\mu_c}{\sigma_c^{2}},
\qquad
\nabla_{\vx}\log p_u(\vx) = -\frac{\vx-\mu_u}{\sigma_u^{2}} .
\]
\smallskip
\noindent\textbf{Step 1: Guided score.}
Insert these into \cref{eq.score_cfg} (CFG):
\begin{align}
\nabla_{\vx}\log\tilde{p}_c(\vx)
&= (1+\omega)\Bigl(-\frac{\vx-\mu_c}{\sigma_c^{2}}\Bigr)
\;+\;
\omega\Bigl(\frac{\vx-\mu_u}{\sigma_u^{2}}\Bigr) \notag\\
&= -\Bigl[\Bigl(\tfrac{1+\omega}{\sigma_c^{2}} - \tfrac{\omega}{\sigma_u^{2}}\Bigr)\vx
\;-\;
\Bigl(\tfrac{1+\omega}{\sigma_c^{2}}\mu_c - \tfrac{\omega}{\sigma_u^{2}}\mu_u\Bigr)
\Bigr].
\label{eq:guided_score_linear}
\end{align}
\smallskip
\noindent\textbf{Step 2: Match to a Gaussian form.}
Any Gaussian $\mathcal{N}(\tilde{\mu}_c,\tilde{\sigma}_c^{2}I)$ has score
$-(\vx-\tilde{\mu}_c)/\tilde{\sigma}_c^{2}$.
Equating with~\cref{eq:guided_score_linear} gives, for all $\vx$,
\begin{align}
\frac{1}{\tilde{\sigma}_c^{2}}
&= \frac{1+\omega}{\sigma_c^{2}} - \frac{\omega}{\sigma_u^{2}}, \label{eq:sigma_coeff}\\
\frac{\tilde{\mu}_c}{\tilde{\sigma}_c^{2}}
&= \frac{1+\omega}{\sigma_c^{2}}\mu_c - \frac{\omega}{\sigma_u^{2}}\mu_u .
\label{eq:mu_coeff}
\end{align}
\smallskip
\noindent\textbf{Step 3: Solve for $\tilde{\sigma}_c$.}
Let $s := \sigma_c^{2}/\sigma_u^{2}\!(>0)$.
Rewrite~\cref{eq:sigma_coeff}:
\[
\frac{1}{\tilde{\sigma}_c^{2}}
= \frac{(1+\omega)-\omega s}{s\,\sigma_u^{2}}
\;\;\Longrightarrow\;\;
\tilde{\sigma}_c^{2}
= \frac{s\,\sigma_u^{2}}{(1+\omega)-\omega s}
= \frac{\sigma_c^{2}}{1+\omega-\omega s},
\]
so that
\[
\boxed{\;
\tilde{\sigma}_c
= \frac{\sigma_c}{\sqrt{\,1+\omega-\omega s\,}}\; }.
\]
\smallskip
\noindent\textbf{Step 4: Solve for $\tilde{\mu}_c$.}
Multiplying~\cref{eq:mu_coeff} by $\tilde{\sigma}_c^{2}$ and
substituting the expression above yields
\[
\tilde{\mu}_c
= \frac{(1+\omega)\mu_c - \omega s\,\mu_u}{1+\omega-\omega s}
= \mu_c + \frac{\omega s}{1+\omega-\omega s}\,(\mu_c-\mu_u).
\]
\end{proof}
\paragraph{Additional Discussion.}
\begin{itemize}[leftmargin=*]
\item \textbf{Consistency with standard CFG.}
When the two Gaussians share the same variance
($\sigma_c=\sigma_u\!\;\Longrightarrow\;s=1$),
\cref{eq.proposed_cfg} reduces to
$\tilde{\sigma}_c=\sigma_c$ and
$\tilde{\mu}_c=\mu_c+\omega(\mu_c-\mu_u)$,
exactly matching the conventional CFG
used in diffusion models~\citep{ho2021classifier}.
\item \textbf{Numerical stability.}
The denominator $1+\omega-\omega s$ can approach $0$ or even
become negative when $s$ is large, causing
$\tilde{\sigma}_c^{2}$ to blow up or change sign.
Intuitively, guidance should \emph{sharpen} $p_c$,
which entails $\tilde{\sigma}_c^{2}\!\le\!\sigma_c^{2}$, i.e.\
$1+\omega-\omega s \ge 1$.
We therefore clip the variance ratio\footnote{This is equivalent to clip the unconditional variance $\sigma_u$ when it is smaller than $\sigma_c$.} to
\[
s \;=\; \textsc{clip}(s,\,0,\,1),
\]
guaranteeing $1+\omega-\omega s\ge1$ for any $\omega>0$
and ensuring both numerical stability and a genuinely
mode-seeking guided distribution.
\end{itemize}
\begin{table}[t]
\centering
\caption{\footnotesize GenEval comparison across different methods.}
\label{tab:method_comparison}
% Resize entire tabular to \textwidth
\resizebox{\textwidth}{!}{%
\begin{tabular}{lccccccc}
\toprule
\textbf{Method} & \textbf{Overall} & \textbf{Single Obj.} & \textbf{Two Obj.} & \textbf{Counting} & \textbf{Colors} & \textbf{Position} & \textbf{Color Attri.} \\
\midrule
\multicolumn{8}{l}{\textbf{Diffusion Models}} \\
SDv1.5~\citep{rombach2022high} & 0.43 & 0.97 & 0.38 & 0.35 & 0.76 & 0.04 & 0.06 \\
PixArt‐$\alpha$~\citep{chen2023pixart} & 0.48 & 0.98 & 0.50 & 0.44 & 0.81 & 0.08 & 0.07 \\
SDv2.1~\citep{rombach2022high} & 0.50 & 0.98 & 0.51 & 0.44 & 0.85 & 0.07 & 0.17 \\
DALL‐E 2~\citep{ramesh2022hierarchical} & 0.52 & 0.94 & 0.66 & 0.49 & 0.77 & 0.10 & 0.19 \\
SDXL~\citep{podell2023sdxl} & \underline{0.55} & 0.98 & 0.74 & 0.39 & 0.85 & 0.15 & 0.23 \\
DALL‐E 3~\citep{betker2023improving} & 0.67 & 0.96 & 0.87 & 0.47 & 0.83 & 0.43 & 0.45 \\
SD3~\citep{esser2024scaling} & 0.74 & 0.99 & 0.94 & 0.72 & 0.89 & 0.33 & 0.60 \\
\addlinespace
\multicolumn{8}{l}{\textbf{Autoregressive Models}} \\
LlamaGen~\citep{sun2024autoregressive} & 0.32 & 0.71 & 0.34 & 0.21 & 0.58 & 0.07 & 0.04 \\
Chameleon~\citep{chameleonteam2024chameleon} & 0.39 & -- & -- & -- & -- & -- & -- \\
Show‐o~\citep{xie2024show} & 0.53 & 0.95 & 0.52 & 0.49 & 0.82 & 0.11 & 0.28 \\
Emu3~\citep{wang2024emu3} & 0.54 & 0.98 & 0.71 & 0.34 & 0.81 & 0.17 & 0.21 \\
\addlinespace
\multicolumn{8}{l}{\textbf{Normalizing Flows}} \\
\model{} (Ours) & 0.56 & 0.97 & 0.58 & 0.47 & 0.77 & 0.20 & 0.34\\
\bottomrule
\end{tabular}%
} % end \resizebox
\end{table}
\begin{thebibliography}{69}
\providecommand{\natexlab}[1]{#1}
\providecommand{\url}[1]{\texttt{#1}}
\expandafter\ifx\csname urlstyle\endcsname\relax
\providecommand{\doi}[1]{doi: #1}\else
\providecommand{\doi}{doi: \begingroup \urlstyle{rm}\Url}\fi
\bibitem[Balaji et~al.(2022)Balaji, Nah, Huang, Vahdat, Song, Kreis, Aittala, Aila, Laine, Catanzaro, et~al.]{balaji2022ediffi}
Balaji, Y., Nah, S., Huang, X., Vahdat, A., Song, J., Kreis, K., Aittala, M., Aila, T., Laine, S., Catanzaro, B., et~al.
\newblock ediffi: Text-to-image diffusion models with an ensemble of expert denoisers.
\newblock \emph{arXiv preprint arXiv:2211.01324}, 2022.
\bibitem[Betker et~al.(2023)Betker, Goh, Jing, Brooks, Wang, Li, Ouyang, Zhuang, Lee, Guo, et~al.]{betker2023improving}
Betker, J., Goh, G., Jing, L., Brooks, T., Wang, J., Li, L., Ouyang, L., Zhuang, J., Lee, J., Guo, Y., et~al.
\newblock Improving image generation with better captions.
\newblock \emph{Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf}, 2\penalty0 (3):\penalty0 8, 2023.
\bibitem[Brown et~al.(2020)Brown, Mann, Ryder, Subbiah, Kaplan, Dhariwal, Neelakantan, Shyam, Sastry, Askell, Agarwal, Herbert-Voss, Krueger, Henighan, Child, Ramesh, Ziegler, Wu, Winter, Hesse, Chen, Sigler, Litwin, Gray, Chess, Clark, Berner, McCandlish, Radford, Sutskever, and Amodei]{gpt3}
Brown, T.~B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.~M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A., Sutskever, I., and Amodei, D.
\newblock Language models are few-shot learners.
\newblock \emph{Advances in neural information processing systems}, 2020.
\bibitem[Chang et~al.(2023)Chang, Zhang, Barber, Maschinot, Lezama, Jiang, Yang, Murphy, Freeman, Rubinstein, Li, and Krishnan]{chang2023muse}
Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.-H., Murphy, K.~P., Freeman, W.~T., Rubinstein, M., Li, Y., and Krishnan, D.
\newblock Muse: Text-to-image generation via masked generative transformers.
\newblock In Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., and Scarlett, J. (eds.), \emph{Proceedings of the 40th International Conference on Machine Learning}, volume 202 of \emph{Proceedings of Machine Learning Research}, pp.\ 4055--4075. PMLR, 23--29 Jul 2023.
\newblock URL \url{https://proceedings.mlr.press/v202/chang23b.html}.
\bibitem[Changpinyo et~al.(2021)Changpinyo, Sharma, Ding, and Soricut]{changpinyo2021cc12m}
Changpinyo, S., Sharma, P., Ding, N., and Soricut, R.
\newblock {Conceptual 12M}: Pushing web-scale image-text pre-training to recognize long-tail visual concepts.
\newblock In \emph{CVPR}, 2021.
\bibitem[Chen et~al.(2023)Chen, Yu, Ge, Yao, Xie, Wu, Wang, Kwok, Luo, Lu, et~al.]{chen2023pixart}
Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wu, Y., Wang, Z., Kwok, J., Luo, P., Lu, H., et~al.
\newblock Pixart-alpha: Fast training of diffusion transformer for photorealistic text-to-image synthesis.
\newblock \emph{arXiv preprint arXiv:2310.00426}, 2023.
\bibitem[Chen et~al.(2018)Chen, Rubanova, Bettencourt, and Duvenaud]{chen2018neural}
Chen, R.~T., Rubanova, Y., Bettencourt, J., and Duvenaud, D.~K.
\newblock Neural ordinary differential equations.
\newblock \emph{Advances in neural information processing systems}, 31, 2018.
\bibitem[Chung et~al.(2022)Chung, Hou, Longpre, Zoph, Tay, Fedus, Li, Wang, Dehghani, Brahma, Webson, Gu, Dai, Suzgun, Chen, Chowdhery, Narang, Mishra, Yu, Zhao, Huang, Dai, Yu, Petrov, Chi, Dean, Devlin, Roberts, Zhou, Le, and Wei]{t5xl}
Chung, H.~W., Hou, L., Longpre, S., Zoph, B., Tay, Y., Fedus, W., Li, E., Wang, X., Dehghani, M., Brahma, S., Webson, A., Gu, S.~S., Dai, Z., Suzgun, M., Chen, X., Chowdhery, A., Narang, S., Mishra, G., Yu, A., Zhao, V., Huang, Y., Dai, A., Yu, H., Petrov, S., Chi, E.~H., Dean, J., Devlin, J., Roberts, A., Zhou, D., Le, Q.~V., and Wei, J.
\newblock Scaling instruction-finetuned language models, 2022.
\newblock URL \url{https://arxiv.org/abs/2210.11416}.
\bibitem[Deng et~al.(2009)Deng, Dong, Socher, Li, Li, and Fei-Fei]{Deng2009ImageNet:Database}
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L.
\newblock {ImageNet: A Large-scale Hierarchical Image Database}.
\newblock \emph{IEEE Conference on Computer Vision and Pattern Recognition}, pp.\ 248--255, 2009.
\bibitem[Dhariwal \& Nichol(2021)Dhariwal and Nichol]{dhariwal2021diffusion}
Dhariwal, P. and Nichol, A.
\newblock Diffusion models beat gans on image synthesis.
\newblock \emph{Advances in Neural Information Processing Systems}, 34:\penalty0 8780--8794, 2021.
\bibitem[Ding et~al.(2021)Ding, Yang, Hong, Zheng, Zhou, Yin, Lin, Zou, Shao, Yang, et~al.]{ding2021cogview}
Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H., et~al.
\newblock Cogview: Mastering text-to-image generation via transformers.
\newblock \emph{Advances in Neural Information Processing Systems}, 34:\penalty0 19822--19835, 2021.
\bibitem[Dinh et~al.(2014)Dinh, Krueger, and Bengio]{dinh2014nice}
Dinh, L., Krueger, D., and Bengio, Y.
\newblock Nice: Non-linear independent components estimation.
\newblock \emph{arXiv preprint arXiv:1410.8516}, 2014.
\bibitem[Dinh et~al.(2016)Dinh, Sohl-Dickstein, and Bengio]{dinh2016density}
Dinh, L., Sohl-Dickstein, J., and Bengio, S.
\newblock Density estimation using real nvp.
\newblock \emph{arXiv preprint arXiv:1605.08803}, 2016.
\bibitem[Dubey et~al.(2024)Dubey, Jauhri, Pandey, Kadian, Al-Dahle, Letman, Mathur, Schelten, Yang, Fan, et~al.]{dubey2024llama}
Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et~al.
\newblock The llama 3 herd of models.
\newblock \emph{arXiv preprint arXiv:2407.21783}, 2024.
\bibitem[Efron(2011)]{efron2011tweedie}
Efron, B.
\newblock Tweedie’s formula and selection bias.
\newblock \emph{Journal of the American Statistical Association}, 106\penalty0 (496):\penalty0 1602--1614, 2011.
\bibitem[Esser et~al.(2021)Esser, Rombach, and Ommer]{taming}
Esser, P., Rombach, R., and Ommer, B.
\newblock Taming transformers for high-resolution image synthesis.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pp.\ 12873--12883, 2021.
\bibitem[Esser et~al.(2024)Esser, Kulal, Blattmann, Entezari, M{\"u}ller, Saini, Levi, Lorenz, Sauer, Boesel, et~al.]{esser2024scaling}
Esser, P., Kulal, S., Blattmann, A., Entezari, R., M{\"u}ller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et~al.
\newblock Scaling rectified flow transformers for high-resolution image synthesis.
\newblock In \emph{Forty-first International Conference on Machine Learning}, 2024.
\bibitem[Fan et~al.(2024)Fan, Li, Qin, Li, Sun, Rubinstein, Sun, He, and Tian]{fan2024fluid}
Fan, L., Li, T., Qin, S., Li, Y., Sun, C., Rubinstein, M., Sun, D., He, K., and Tian, Y.
\newblock Fluid: Scaling autoregressive text-to-image generative models with continuous tokens.
\newblock \emph{arXiv preprint arXiv:2410.13863}, 2024.
\bibitem[Gafni et~al.(2022)Gafni, Polyak, Ashual, Sheynin, Parikh, and Taigman]{makeascene}
Gafni, O., Polyak, A., Ashual, O., Sheynin, S., Parikh, D., and Taigman, Y.
\newblock Make-a-scene: Scene-based text-to-image generation with human priors.
\newblock 2022.
\newblock \doi{10.48550/ARXIV.2203.13131}.
\newblock URL \url{https://arxiv.org/abs/2203.13131}.
\bibitem[Ghosh et~al.(2023)Ghosh, Hajishirzi, and Schmidt]{ghosh2023geneval}
Ghosh, D., Hajishirzi, H., and Schmidt, L.
\newblock Geneval: An object-focused framework for evaluating text-to-image alignment.
\newblock \emph{Advances in Neural Information Processing Systems}, 36:\penalty0 52132--52152, 2023.
\bibitem[Goodfellow et~al.(2016)Goodfellow, Bengio, and Courville]{Goodfellow-et-al-2016}
Goodfellow, I., Bengio, Y., and Courville, A.
\newblock \emph{Deep Learning}.
\newblock MIT Press, 2016.
\newblock \url{http://www.deeplearningbook.org}.
\bibitem[Grathwohl et~al.(2018)Grathwohl, Chen, Bettencourt, Sutskever, and Duvenaud]{grathwohl2018ffjord}
Grathwohl, W., Chen, R.~T., Bettencourt, J., Sutskever, I., and Duvenaud, D.
\newblock Ffjord: Free-form continuous dynamics for scalable reversible generative models.
\newblock \emph{arXiv preprint arXiv:1810.01367}, 2018.
\bibitem[Gu et~al.(2024{\natexlab{a}})Gu, Shen, Zhai, Zhang, Jaitly, and Susskind]{gu2024kaleido}
Gu, J., Shen, Y., Zhai, S., Zhang, Y., Jaitly, N., and Susskind, J.~M.
\newblock Kaleido diffusion: Improving conditional diffusion models with autoregressive latent modeling.
\newblock \emph{arXiv preprint arXiv:2405.21048}, 2024{\natexlab{a}}.
\bibitem[Gu et~al.(2024{\natexlab{b}})Gu, Wang, Zhang, Zhang, Zhang, Jaitly, Susskind, and Zhai]{gu2024dart}
Gu, J., Wang, Y., Zhang, Y., Zhang, Q., Zhang, D., Jaitly, N., Susskind, J., and Zhai, S.
\newblock Dart: Denoising autoregressive transformer for scalable text-to-image generation.
\newblock \emph{arXiv preprint arXiv:2410.08159}, 2024{\natexlab{b}}.
\bibitem[Guo et~al.(2025)Guo, Yang, Zhang, Song, Zhang, Xu, Zhu, Ma, Wang, Bi, et~al.]{guo2025deepseek}
Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et~al.
\newblock Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.
\newblock \emph{arXiv preprint arXiv:2501.12948}, 2025.
\bibitem[Heusel et~al.(2017)Heusel, Ramsauer, Unterthiner, Nessler, and Hochreiter]{fid}
Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S.
\newblock Gans trained by a two time-scale update rule converge to a local nash equilibrium.
\newblock \emph{Advances in neural information processing systems}, 30, 2017.
\bibitem[Ho \& Salimans(2021)Ho and Salimans]{ho2021classifier}
Ho, J. and Salimans, T.
\newblock Classifier-free diffusion guidance.
\newblock In \emph{NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications}, 2021.
\bibitem[Ho et~al.(2019)Ho, Chen, Srinivas, Duan, and Abbeel]{ho2019flow++}
Ho, J., Chen, X., Srinivas, A., Duan, Y., and Abbeel, P.
\newblock Flow++: Improving flow-based generative models with variational dequantization and architecture design.
\newblock In \emph{International conference on machine learning}, pp.\ 2722--2730. PMLR, 2019.
\bibitem[Ho et~al.(2020)Ho, Jain, and Abbeel]{ho2020denoising}
Ho, J., Jain, A., and Abbeel, P.
\newblock Denoising diffusion probabilistic models.
\newblock \emph{Advances in Neural Information Processing Systems}, 33:\penalty0 6840--6851, 2020.
\bibitem[Ho et~al.(2022{\natexlab{a}})Ho, Chan, Saharia, Whang, Gao, Gritsenko, Kingma, Poole, Norouzi, Fleet, et~al.]{ho2022imagen}
Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D.~P., Poole, B., Norouzi, M., Fleet, D.~J., et~al.
\newblock Imagen video: High definition video generation with diffusion models.
\newblock \emph{arXiv preprint arXiv:2210.02303}, 2022{\natexlab{a}}.
\bibitem[Ho et~al.(2022{\natexlab{b}})Ho, Saharia, Chan, Fleet, Norouzi, and Salimans]{ho2022cascaded}
Ho, J., Saharia, C., Chan, W., Fleet, D.~J., Norouzi, M., and Salimans, T.
\newblock Cascaded diffusion models for high fidelity image generation.
\newblock \emph{J. Mach. Learn. Res.}, 23:\penalty0 47--1, 2022{\natexlab{b}}.
\bibitem[Hutchinson(1989)]{hutchinson1989stochastic}
Hutchinson, M.~F.
\newblock A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines.
\newblock \emph{Communications in Statistics-Simulation and Computation}, 18\penalty0 (3):\penalty0 1059--1076, 1989.
\bibitem[Jabri et~al.(2022)Jabri, Fleet, and Chen]{jabri2022scalable}
Jabri, A., Fleet, D., and Chen, T.
\newblock Scalable adaptive computation for iterative generation.
\newblock \emph{arXiv preprint arXiv:2212.11972}, 2022.
\bibitem[Kang et~al.(2023)Kang, Zhu, Zhang, Park, Shechtman, Paris, and Park]{kang2023scaling}
Kang, M., Zhu, J.-Y., Zhang, R., Park, J., Shechtman, E., Paris, S., and Park, T.
\newblock Scaling up gans for text-to-image synthesis.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pp.\ 10124--10134, 2023.
\bibitem[Kaplan et~al.(2020)Kaplan, McCandlish, Henighan, Brown, Chess, Child, Gray, Radford, Wu, and Amodei]{kaplan2020scaling}
Kaplan, J., McCandlish, S., Henighan, T., Brown, T.~B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J., and Amodei, D.
\newblock Scaling laws for neural language models.
\newblock \emph{arXiv preprint arXiv:2001.08361}, 2020.
\bibitem[Karras et~al.(2024)Karras, Aittala, Kynk{\"a}{\"a}nniemi, Lehtinen, Aila, and Laine]{karras2024guiding}
Karras, T., Aittala, M., Kynk{\"a}{\"a}nniemi, T., Lehtinen, J., Aila, T., and Laine, S.
\newblock Guiding a diffusion model with a bad version of itself.
\newblock \emph{Advances in Neural Information Processing Systems}, 37:\penalty0 52996--53021, 2024.
\bibitem[Kingma \& Welling(2013)Kingma and Welling]{kingma2013auto}
Kingma, D.~P. and Welling, M.
\newblock Auto-encoding variational bayes.
\newblock \emph{arXiv preprint arXiv:1312.6114}, 2013.
\bibitem[Kingma et~al.(2016)Kingma, Salimans, Jozefowicz, Chen, Sutskever, and Welling]{kingma2016improved}
Kingma, D.~P., Salimans, T., Jozefowicz, R., Chen, X., Sutskever, I., and Welling, M.
\newblock Improved variational inference with inverse autoregressive flow.
\newblock \emph{Advances in neural information processing systems}, 29, 2016.
\bibitem[Lee et~al.(2022)Lee, Kim, Kim, Cho, and Han]{lee2022autoregressive}
Lee, D., Kim, C., Kim, S., Cho, M., and Han, W.-S.
\newblock Autoregressive image generation using residual quantization.
\newblock In \emph{Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pp.\ 11523--11532, 2022.
\bibitem[Li et~al.(2024)Li, Tian, Li, Deng, and He]{li2024autoregressive}
Li, T., Tian, Y., Li, H., Deng, M., and He, K.
\newblock Autoregressive image generation without vector quantization.
\newblock \emph{arXiv preprint arXiv:2406.11838}, 2024.
\bibitem[Li et~al.(2025)Li, Sun, Fan, and He]{li2025fractal}
Li, T., Sun, Q., Fan, L., and He, K.
\newblock Fractal generative models.
\newblock \emph{arXiv preprint arXiv:2502.17437}, 2025.
\bibitem[Liang et~al.(2024)Liang, Xu, Hong, Shang, Wang, Fu, and Liu]{liang2024survey}
Liang, Z., Xu, Y., Hong, Y., Shang, P., Wang, Q., Fu, Q., and Liu, K.
\newblock A survey of multimodel large language models.
\newblock In \emph{Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering}, pp.\ 405--409, 2024.
\bibitem[Lin et~al.(2014)Lin, Maire, Belongie, Hays, Perona, Ramanan, Doll{\'{a}}r, and Zitnick]{Lin2014}
Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll{\'{a}}r, P., and Zitnick, C.~L.
\newblock {Microsoft COCO: Common Objects in Context}.
\newblock \emph{European Conference on Computer Vision}, pp.\ 740--755, 2014.
\bibitem[Lipman et~al.(2023)Lipman, Chen, Ben-Hamu, Nickel, and Le]{lipman2023flow}
Lipman, Y., Chen, R. T.~Q., Ben-Hamu, H., Nickel, M., and Le, M.
\newblock Flow matching for generative modeling.
\newblock In \emph{The Eleventh International Conference on Learning Representations}, 2023.
\newblock URL \url{https://openreview.net/forum?id=PqvMRDCJT9t}.
\bibitem[Ma et~al.(2024)Ma, Goldstein, Albergo, Boffi, Vanden-Eijnden, and Xie]{ma2024sit}
Ma, N., Goldstein, M., Albergo, M.~S., Boffi, N.~M., Vanden-Eijnden, E., and Xie, S.
\newblock Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers.
\newblock \emph{arXiv preprint arXiv:2401.08740}, 2024.
\bibitem[OpenAI(2024)]{openai2024gpt4o}
OpenAI.
\newblock Gpt‑4o system card.
\newblock \url{https://openai.com/index/gpt-4o-system-card/}, 2024.
\newblock Accessed: April 12, 2025.
\bibitem[Papamakarios et~al.(2017)Papamakarios, Murray, and Pavlakou]{papamakarios2017masked}
Papamakarios, G., Murray, I., and Pavlakou, T.
\newblock Masked autoregressive flow for density estimation.
\newblock In Guyon, I., von Luxburg, U., Bengio, S., Wallach, H.~M., Fergus, R., Vishwanathan, S. V.~N., and Garnett, R. (eds.), \emph{Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, {USA}}, pp.\ 2338--2347, 2017.
\newblock URL \url{https://proceedings.neurips.cc/paper/2017/hash/6c1da886822c67822bcf3679d04369fa-Abstract.html}.
\bibitem[Peebles \& Xie(2023)Peebles and Xie]{peebles2023scalable}
Peebles, W. and Xie, S.
\newblock Scalable diffusion models with transformers.
\newblock In \emph{Proceedings of the IEEE/CVF International Conference on Computer Vision}, pp.\ 4195--4205, 2023.
\bibitem[Podell et~al.(2023)Podell, English, Lacey, Blattmann, Dockhorn, M{\"u}ller, Penna, and Rombach]{podell2023sdxl}
Podell, D., English, Z., Lacey, K., Blattmann, A., Dockhorn, T., M{\"u}ller, J., Penna, J., and Rombach, R.
\newblock Sdxl: improving latent diffusion models for high-resolution image synthesis.
\newblock \emph{arXiv preprint arXiv:2307.01952}, 2023.
\bibitem[Raffel et~al.(2020)Raffel, Shazeer, Roberts, Lee, Narang, Matena, Zhou, Li, and Liu]{raffel2020exploring}
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., and Liu, P.~J.
\newblock Exploring the limits of transfer learning with a unified text-to-text transformer.
\newblock \emph{Journal of machine learning research}, 21\penalty0 (140):\penalty0 1--67, 2020.
\bibitem[Ramesh et~al.(2021)Ramesh, Pavlov, Goh, Gray, Voss, Radford, Chen, and Sutskever]{ramesh2021zero}
Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., and Sutskever, I.
\newblock Zero-shot text-to-image generation.
\newblock In \emph{International Conference on Machine Learning}, pp.\ 8821--8831. PMLR, 2021.
\bibitem[Ramesh et~al.(2022)Ramesh, Dhariwal, Nichol, Chu, and Chen]{ramesh2022hierarchical}
Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M.
\newblock Hierarchical text-conditional image generation with clip latents.
\newblock \emph{arXiv preprint arXiv:2204.06125}, 2022.
\bibitem[Rezende \& Mohamed(2015)Rezende and Mohamed]{pmlr-v37-rezende15}
Rezende, D. and Mohamed, S.
\newblock Variational inference with normalizing flows.
\newblock In Bach, F. and Blei, D. (eds.), \emph{Proceedings of the 32nd International Conference on Machine Learning}, volume~37 of \emph{Proceedings of Machine Learning Research}, pp.\ 1530--1538, Lille, France, 07--09 Jul 2015. PMLR.
\newblock URL \url{https://proceedings.mlr.press/v37/rezende15.html}.
\bibitem[Rombach et~al.(2022)Rombach, Blattmann, Lorenz, Esser, and Ommer]{rombach2022high}
Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B.
\newblock High-resolution image synthesis with latent diffusion models.
\newblock In \emph{Proceedings of the IEEE/CVF conference on computer vision and pattern recognition}, pp.\ 10684--10695, 2022.
\bibitem[Su et~al.(2024)Su, Ahmed, Lu, Pan, Bo, and Liu]{su2024roformer}
Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., and Liu, Y.
\newblock Roformer: Enhanced transformer with rotary position embedding.
\newblock \emph{Neurocomputing}, 568:\penalty0 127063, 2024.
\bibitem[Sun et~al.(2024)Sun, Jiang, Chen, Zhang, Peng, Luo, and Yuan]{sun2024autoregressive}
Sun, P., Jiang, Y., Chen, S., Zhang, S., Peng, B., Luo, P., and Yuan, Z.
\newblock Autoregressive model beats diffusion: Llama for scalable image generation.
\newblock \emph{arXiv preprint arXiv:2406.06525}, 2024.
\bibitem[Team(2024)]{chameleonteam2024chameleon}
Team, C.
\newblock Chameleon: Mixed-modal early-fusion foundation models, 2024.
\bibitem[Team et~al.(2024)Team, Riviere, Pathak, Sessa, Hardin, Bhupatiraju, Hussenot, Mesnard, Shahriari, Ram{\'e}, et~al.]{team2024gemma}
Team, G., Riviere, M., Pathak, S., Sessa, P.~G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram{\'e}, A., et~al.
\newblock Gemma 2: Improving open language models at a practical size.
\newblock \emph{arXiv preprint arXiv:2408.00118}, 2024.
\bibitem[Tian et~al.(2024)Tian, Jiang, Yuan, Peng, and Wang]{tian2024visual}
Tian, K., Jiang, Y., Yuan, Z., Peng, B., and Wang, L.
\newblock Visual autoregressive modeling: Scalable image generation via next-scale prediction.
\newblock \emph{arXiv preprint arXiv:2404.02905}, 2024.
\bibitem[Tschannen et~al.(2024{\natexlab{a}})Tschannen, Eastwood, and Mentzer]{tschannen2024givt}
Tschannen, M., Eastwood, C., and Mentzer, F.
\newblock Givt: Generative infinite-vocabulary transformers.
\newblock In \emph{European Conference on Computer Vision}, pp.\ 292--309. Springer, 2024{\natexlab{a}}.
\bibitem[Tschannen et~al.(2024{\natexlab{b}})Tschannen, Pinto, and Kolesnikov]{tschannen2024jetformer}
Tschannen, M., Pinto, A.~S., and Kolesnikov, A.
\newblock Jetformer: An autoregressive generative model of raw images and text.
\newblock \emph{arXiv preprint arXiv:2411.19722}, 2024{\natexlab{b}}.
\bibitem[Vaswani et~al.(2017)Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin]{transformer}
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.~N., Kaiser, L., and Polosukhin, I.
\newblock {Attention Is All You Need}.
\newblock \emph{Advances in Neural Information Processing Systems}, pp.\ 5998--6008, 2017.
\bibitem[Wang et~al.(2024)Wang, Zhang, Luo, Sun, Cui, Wang, Zhang, Wang, Li, Yu, et~al.]{wang2024emu3}
Wang, X., Zhang, X., Luo, Z., Sun, Q., Cui, Y., Wang, J., Zhang, F., Wang, Y., Li, Z., Yu, Q., et~al.
\newblock Emu3: Next-token prediction is all you need.
\newblock \emph{arXiv preprint arXiv:2409.18869}, 2024.
\bibitem[Xie et~al.(2024)Xie, Mao, Bai, Zhang, Wang, Lin, Gu, Chen, Yang, and Shou]{xie2024show}
Xie, J., Mao, W., Bai, Z., Zhang, D.~J., Wang, W., Lin, K.~Q., Gu, Y., Chen, Z., Yang, Z., and Shou, M.~Z.
\newblock Show-o: One single transformer to unify multimodal understanding and generation.
\newblock \emph{arXiv preprint arXiv:2408.12528}, 2024.
\bibitem[Yu et~al.(2022)Yu, Xu, Koh, Luong, Baid, Wang, Vasudevan, Ku, Yang, Ayan, et~al.]{yu2022scaling}
Yu, J., Xu, Y., Koh, J.~Y., Luong, T., Baid, G., Wang, Z., Vasudevan, V., Ku, A., Yang, Y., Ayan, B.~K., et~al.
\newblock Scaling autoregressive models for content-rich text-to-image generation.
\newblock \emph{Transactions on Machine Learning Research}, 2022.
\bibitem[Zhai et~al.(2024)Zhai, Zhang, Nakkiran, Berthelot, Gu, Zheng, Chen, Bautista, Jaitly, and Susskind]{zhai2024normalizing}
Zhai, S., Zhang, R., Nakkiran, P., Berthelot, D., Gu, J., Zheng, H., Chen, T., Bautista, M.~A., Jaitly, N., and Susskind, J.
\newblock Normalizing flows are capable generative models.
\newblock \emph{arXiv preprint arXiv:2412.06329}, 2024.
\bibitem[Zheng et~al.(2024{\natexlab{a}})Zheng, Nie, Vahdat, and Anandkumar]{Zheng2024MaskDiT}
Zheng, H., Nie, W., Vahdat, A., and Anandkumar, A.
\newblock Fast training of diffusion models with masked transformers.
\newblock In \emph{Transactions on Machine Learning Research (TMLR)}, 2024{\natexlab{a}}.
\bibitem[Zheng et~al.(2024{\natexlab{b}})Zheng, Wang, Yuan, Ning, He, You, Yang, and Zhou]{zheng2024learning}
Zheng, H., Wang, Z., Yuan, J., Ning, G., He, P., You, Q., Yang, H., and Zhou, M.
\newblock Learning stackable and skippable {LEGO} bricks for efficient, reconfigurable, and variable-resolution diffusion modeling.
\newblock In \emph{The Twelfth International Conference on Learning Representations}, 2024{\natexlab{b}}.
\newblock URL \url{https://openreview.net/forum?id=qmXedvwrT1}.
\bibitem[Zhou et~al.(2024)Zhou, Yu, Babu, Tirumala, Yasunaga, Shamis, Kahn, Ma, Zettlemoyer, and Levy]{zhou2024transfusion}
Zhou, C., Yu, L., Babu, A., Tirumala, K., Yasunaga, M., Shamis, L., Kahn, J., Ma, X., Zettlemoyer, L., and Levy, O.
\newblock Transfusion: Predict the next token and diffuse images with one multi-modal model.
\newblock \emph{arXiv preprint arXiv:2408.11039}, 2024.
\end{thebibliography}