FreeAnimate: Training-Free Human Image Animation with Preview-Guided Denoising
Extracted source text · arXiv 2606.06885
% Template for ICASSP-2026 paper; to be used with:
% spconf.sty - ICASSP/ICIP LaTeX style file, and
% IEEEbib.bst - IEEE bibliography style file.
% --------------------------------------------------------------------------
\documentclass{article}
\usepackage{spconf,amsmath,graphicx,hyperref}
\usepackage{lipsum}
\usepackage{amssymb}
\usepackage{booktabs}
\usepackage{xcolor}
\definecolor{darkgreen}{rgb}{0.0, 0.5, 0.0}
\usepackage[table]{xcolor}
\usepackage{makecell}
\usepackage{multirow}
\usepackage{graphicx}
\usepackage{todonotes}
\usepackage{verbatim}
% Example definitions.
% --------------------
\def\x{{\mathbf x}}
\def\L{{\cal L}}
% Title.
% ------
% \title{EgoDiff-Pressure: Conditional Video Diffusion for Continuous Hand Pressure Estimation from Egocentric Views}
\title{FreeAnimate: Training-Free Human Image Animation with Preview-Guided Denoising}
% EgoDiff-Pressure: Conditional Video Diffusion for Continuous Hand Pressure Estimation from Egocentric Views
%
% Single address.
% ---------------
\name{Yuan Zeng\textsuperscript{1} \qquad Yujia Shi\textsuperscript{2,3} \qquad Zongqing Lu\textsuperscript{1} \qquad QingMin Liao\textsuperscript{1}$^{\dagger}$
\thanks{ $\dagger$ Corresponding author.}}
\address{\textsuperscript{1}Shenzhen International Graduate School, Tsinghua University, China\\
\textsuperscript{2}Harbin Institute of Technology, China\\
\textsuperscript{3}Pengcheng Laboratory, China}
% \name{Author Name$^{\star \dagger}$ \qquad Author Name$^{\star}$ \qquad Author Name$^{\dagger}$}
% \address{$^{\star}$ Affiliation Number One \\
% $^{\dagger}$ Affiliation Number Two}
%
\begin{document}
\ninept
%
\maketitle
%
\begin{abstract}
Human Image Animation has seen significant advancements, primarily driven by diffusion models. However, existing methods typically demand substantial training data and resources to achieve high-quality results, limiting generalization and accessibility. In this work, we introduce \emph{FreeAnimate}, a training-free framework that leverages the inherent capabilities of image diffusion models to enable temporal consistency, identity preservation, and background stability. Our approach incorporates a novel preview generation strategy that provides temporal and structural priors from generated preview frames, effectively guiding pose alignment and background consistency without training. Additionally, FreeAnimate introduces Inversion-Boosted Attention and Reference-Anchored Self-Attention modules to guarantee temporal consistency and identity preservation. Experimental results demonstrate that FreeAnimate outperforms existing training-free competitors and training-based baseline methods, achieving generation quality comparable to state-of-the-art methods and offering robust generalization across diverse datasets. Our project page is at
\href{https://freeani.github.io/}{https://freeani.github.io/}.
\end{abstract}
\begin{keywords}
Human Video Generation, Training-Free Method, Human Image Animation
\end{keywords}
\begin{figure}[h] \centering
\includegraphics[width=0.45\textwidth]{figures/Fig1_v3.pdf}
\caption{
FreeAnimate uses preview frames for animation, outperforming DisCo's fixed foreground-background guidance and achieving quality on par with Champ's multi-conditioning generation. Each image's left side shows conditioning inputs, with the second row of FreeAnimate's inputs representing the cross attention map of preview frames generated by DDIM inversion.
}
\label{fig:figure1}
\vspace{-3mm}
\end{figure}
\section{Introduction}
\label{sec:intro}
The development of diffusion models has greatly advanced applications in content generation, with human-centered themes like Human Image Animation (HIA) receiving particular attention. HIA \cite{MRAA,TPS,Disco} holds strong potential across various fields, including social media, entertainment industry, and video game development.
Initial attempts with GAN-based HIA \cite{GAN} suffered from unstable training and weak controllability, yielding low-quality results and slowing development of the field.
% Recently, diffusion models \cite{DDPM,DDIM} have shown great potential in image and video generation, gradually supplanting GANs as the preferred base model. The introduction of the ControlNet \cite{ControlNet} has further endowed diffusion models with powerful control capabilities.
% The diffusion-based HIA method, DisCo \cite{Disco}, has significantly outperformed leading GAN-based approaches such as MRAA \cite{MRAA} and TPS \cite{TPS}. Consequently, numerous diffusion-based HIA methods have been proposed.
% MagicAnimate \cite{Magicanimate} and MagicPose \cite{MagicPose}, composed of three key components: a denoising U-Net \cite{U_Net}, an AppearanceNet injecting reference features into U-Net, and a pose ControlNet \cite{ControlNet}.
% To improve generation speed and quality, AnimateAnyone \cite{Animateanyone} and MimicMotion \cite{Mimicmotion} replace the computationally intensive pose ControlNet with a lightweight ConvNet-based module.
% Champ \cite{Champ} leverages the 3D human parametric model SMPL \cite{SMPL} to incorporate additional controls during generation.
% The recently proposed method, StableAnimator \cite{StableAnimator}, introduces a global content-aware Face Encoder and a distribution-aware ID Adapter to integrate face embeddings into video diffusion models, achieving better identity preservation.
Recently, diffusion models \cite{DDPM,DDIM} have become the dominant backbone for image and video generation, gradually replacing GANs. With modules such as ControlNet \cite{ControlNet}, they enable strong controllability, and diffusion-based HIA methods like DisCo \cite{Disco} already surpass leading GAN-based approaches \cite{MRAA,TPS}. Building on this, many variants have been proposed: MagicAnimate and MagicPose \cite{Magicanimate,MagicPose} combine U-Net denoising with AppearanceNet and pose ControlNet; AnimateAnyone and MimicMotion \cite{Animateanyone,Mimicmotion} pursue efficiency by replacing heavy ControlNets with lightweight ConvNets; Champ \cite{Champ} integrates 3D parametric cues from SMPL \cite{SMPL}; and StableAnimator \cite{StableAnimator} further enhances identity fidelity with dedicated face encoders and ID adapters.
% However, all of the above methods require extensive training data and resources to achieve high-quality HIA results, and the generated frames often reflect biases of the training data.
% With the rapid advancement of generative models, increasingly powerful foundation models and image/video editing techniques are emerging. These models inherently possess strong generative capabilities, making it possible to achieve high-quality HIA results without additional training.
% Motivated by this, we propose a training-free HIA framework that leverages existing diffusion-based image generation model, aiming to further inspire research in this emerging area.
% To the best of our knowledge, PoseAnimate \cite{PoseAnimate} was the only training-free method prior to our work. It ensures pose alignment through the Pose-Aware Control Module, and enhances temporal coherence with the Dual Consistency Attention Module. Nevertheless, PoseAnimate mainly showcases cartoon animations with limited examples of animations in real-world scenarios, and its complex architecture hinders researchers seeking to conduct further studies.
However, existing methods still rely on large-scale training and resources, and their results often inherit dataset biases. With the rise of powerful foundation models and editing tools, high-quality HIA can now be pursued without extra training. Motivated by this, we propose a training-free framework that leverages pretrained diffusion models. To our knowledge, the only prior training-free approach is PoseAnimate \cite{PoseAnimate}, which ensures pose alignment and temporal coherence through specialized modules. Yet PoseAnimate mainly demonstrates cartoon examples and employs a complex architecture, limiting its applicability to real-world scenarios and further research.
% To address the aforementioned issues, we introduces \emph{FreeAnimate}, a HIA framework designed to offer training-free capability and enhanced generalization.
% Existing approaches depend on either extensive training datasets or temporal modules \cite{Animatediff} to maintain temporal coherence.
% In contrast, to incorporate temporal consistency priors without additional training, we introduce the Preview Generation Strategy. Leveraging well-pretrained models such as Grounded-SAM \cite{GroundedSAM}, image editing model \cite{Masactrl, T2iAdapter}, and image inpainting model \cite{Mat}, we design a pipeline to generate preview frames that provide temporal consistency priors aligned with the target frames.
To address these issues, we propose \emph{FreeAnimate}. Unlike prior methods that rely on large datasets or temporal modules \cite{Animatediff}, we introduce a \emph{Preview Generation Strategy} that leverages pretrained models such as Grounded-SAM \cite{GroundedSAM}, editing \cite{Masactrl, T2iAdapter}, and inpainting \cite{Mat} to produce preview frames, providing temporal priors aligned with target frames.
% To further enhance generation quality, we design a training-free framework based on image diffusion model \cite{stablediffusion, ControlNet} that works with the preview generation strategy. Unlike existing methods incorporate CLIP image encoders and AppearanceNet, our framework consists only of a pose ControlNet and U-Net for basic pose-guided image synthesis. Additionally, Inversion-Boosted Attention is proposed, leveraging preview frame attention maps from DDIM inversion \cite{DDIM} to guide denoising process, and Reference-Anchored Self-Attention is also integrated to ensure identity preservation by enabling each latent interacts with the reference latent.
To further improve quality, we design a training-free framework built on an image diffusion model \cite{stablediffusion} with the preview strategy. Unlike methods that add CLIP encoders or AppearanceNet, our framework uses only a pose ControlNet and U-Net for pose-guided synthesis. We also introduce \emph{Inversion-Boosted Attention}, which leverages preview-frame attention maps from DDIM inversion \cite{DDIM} to guide denoising, and \emph{Reference-Anchored Self-Attention}, which preserves identity by letting each latent attend to the reference latent.
% As shown in Figure~\ref{fig:figure1}, DisCo guides image animation using pose sequences, alongside disentangled foreground and background. Champ, in contrast, employs depth, normal, semantic, and pose sequences to enhance motion guidance. Our approach utilizes inverted noise and attention maps derived from preview frames closest to the target frames, demonstrating superior results over the well-established baseline Disco and achieving generation quality on par with the state-of-the-art method Champ.
As shown in Figure~\ref{fig:figure1}, our approach utilizes inverted noise and attention maps derived from preview frames closest to the target frames, demonstrating superior results over the well-established baseline DisCo and achieving generation quality on par with the state-of-the-art method Champ.
Our contributions are summarized as follows: (1) We propose FreeAnimate, a simple, training-free framework for human image animation. (2) We introduce the Preview Generation Strategy, a framework-agnostic pipeline that generates high-quality preview frames for human image animation, significantly improving both video fidelity and temporal coherence. (3) Experiments on various benchmarks demonstrate FreeAnimate's robust generalization and high-quality generation performance.
\section{Method}
\label{sec:method}
Given a reference image $ {I_{ref}} \in \mathbb{R}^{H \times W \times 3} $ that contains an identity, and a pose sequence ${p_{1:N}} = [p_1, \dots, p_N] \in \mathbb{R}^{N \times H \times W \times 3} $, where N is the number of frames. Human image animation aims to generate a temporally coherent video ${I_{1:N}} = [I_1, \dots, I_N]$ in which both identity and background appearance follow the reference image, while the identity pose conforms to the given pose sequence.
% In what follows, we first give the model architecture of FreeAnimate in Section \ref{sub:NetworkArchitecture}. After that, Section \ref{sub:PreviewGenerationStrategy} describes the Preview Generation Strategy. For more details on the preliminary information, please refer to the supplementary material. Figure~\ref{fig:figure2} shows the overview of FreeAnimate.
\subsection{Network Architecture}
\label{sub:NetworkArchitecture}
\textbf{Overview.}
Due to the sampling-based nature of diffusion models, generated images may exhibit unexpected artifacts or deformations. While subtle in single images, these inconsistencies become more noticeable in video.
To address this issue, many training-based HIA methods \cite{Animateanyone, Champ, Disco, Mimicmotion, StableAnimator} incorporate reference image features into the U-Net via cross-attention, using embeddings from a fixed CLIP image encoder. However, we observe that directly injecting image features without further training yields suboptimal results.
We attribute this to two primary factors: (1) a domain gap, as the SD model's cross-attention was well-trained with CLIP text embeddings, and (2) semantic-level CLIP image embeddings being too compact to capture fine-grained details.
Some methods \cite{Magicanimate, Champ, FYPv2} use an AppearanceNet, typically a trainable copy of the U-Net, to extract fine-grained reference features. While effective, it adds significant computational overhead.
% We introduce a simple network architecture that combines a U-Net with ControlNet, offering a more efficient way to incorporating reference image content and structure into the U-Net without relying on CLIP or AppearanceNet. Specifically, our method incorporates three components in addition to the basic SD model.
We propose a simple architecture combining U-Net with ControlNet, which integrates reference image content and structure without CLIP or AppearanceNet. Our method incorporates three components to the basic SD model.
\begin{figure}[tb] \centering
\includegraphics[width=0.47\textwidth]{figures/Fig2_v4.pdf}
\caption{The proposed FreeAnimate framework. Given a reference image and pose sequence, FreeAnimate first generates preview frames using Preview Generation Strategy, then applies DDIM inversion with RA-SA to obtain the initial noise, and stores the resulting attention maps. During denoising, these maps are reused in Inversion-Boosted Attention to preserve structural consistency.} \label{fig:figure2}
\vspace{-2mm}
\end{figure}
\noindent \textbf{Control Branch.}
% In HIA task, a basic requirement is that the poses in the generated frames match the given pose sequence. To achieve this, most methods incorporate a control branch as a fundamental module.
% We introduce a control branch consisting of a well-trained, publicly available ControlNet \cite{ControlNet}, which serves as a means to incorporate pose conditions into the U-Net.
% Pose information can also be incorporated using more efficient methods, such as the T2I-Adapter \cite{T2iAdapter}, or by training a custom lightweight ConvNet-based module. As suggested in \cite{ControlNetDiscussion188}, these approaches can effectively replace ControlNet while maintaining comparable controllability and image quality. However, as our goal is to construct a training-free framework, we directly utilize readily available ControlNet to build our model, ensuring the generality and robustness of our approach.
In HIA, a key requirement is that generated poses match the input sequence, typically achieved with a control branch. We adopt a well-trained, publicly available ControlNet \cite{ControlNet} to inject pose conditions into the U-Net. Although alternatives like T2I-Adapter \cite{T2iAdapter} or lightweight ConvNets \cite{ControlNetDiscussion188} can offer similar controllability, we directly use ControlNet to ensure a training-free, general, and robust framework.
\noindent \textbf{Inversion-Boosted Attention (IBA).}
% Using inverted noise as the initial noise for the denoising process is a common technique in video and image editing \cite{Pix2video,Tokenflow,TuneAvideo,VidtoMe}. In our approach, we leverage inverted noise from preview frames to improve temporal consistency and structural alignment.
% However, directly feeding the inverted noise into the denoising U-Net is insufficient, as cumulative errors across $T$ steps of DDIM inversion degrade the original motion and structural details \cite{Fatezero}.
Using inverted noise as the initial noise is common in video and image editing. We adopt it from preview frames to enhance temporal consistency and structure. However, feeding it directly into the U-Net causes cumulative DDIM inversion errors that degrade motion and detail \cite{Fatezero}.
To address this issue, we propose Inversion-Boosted Attention (IBA), which enhances structural consistency by modifying both attention mechanisms.
Inspired by FateZero \cite{Fatezero}, IBA offers an effective solution by utilizing DDIM inversion attention maps from preview frames to guide the denoising process.
As shown in the upper part of Fig.~\ref{fig:figure2}, at each step $t$ of the inversion process, we store the generated self-attention maps $\textit{s}_{t}^{\text{ pre}}$ and cross-attention maps $\textit{c}_{t}^{\text{ pre}}$ from the preview frames $z_{0}^{\text{pre}}$ as follows:
% \vspace{2mm}
\begin{equation}
z_{T},\left[\textit{s}_{t}^{\text{ pre}}\right]_{t=1}^{T},\left[\textit{c}_{t}^{\text{ pre}}\right]_{t=1}^{T}=\operatorname{DDIM-I\scalebox{0.75}{NV}}\left(z_{0}^{\text{pre}}\right),
\end{equation}
where DDIM-I\scalebox{0.75}{NV} denotes the DDIM inversion process.
During the denoising process, the inversion attention maps $\left[\textit{s}_{t}^{\text{ pre}}\right]_{t=1}^{T}$ and $\left[\textit{c}_{t}^{\text{ pre}}\right]_{t=1}^{T}$ are used to compute the attention. Formally,
\begin{equation}
\begin{split}
\text{S\scalebox{0.75}{ELF}-A\scalebox{0.75}{TT}}=\operatorname{\scalebox{0.9}{Softmax}} \left(\frac{Q_{s}^{\text{pre}} {K_{s}^{\text{pre}}}^T}{\sqrt{d}}\right) \cdot V_{s}
=\textit{s}_{t}^{\text{ pre}} \cdot V_{s},
\end{split}
\end{equation}
\begin{equation}
\begin{split}
\text{C\scalebox{0.75}{ROSS}-A\scalebox{0.75}{TT}}=\operatorname{\scalebox{0.9}{Softmax}} \left(\frac{Q_{c}^{\text{pre}} {K_{c}^{\text{pre}}}^T}{\sqrt{d}}\right) \cdot V_{c}
=\textit{c}_{t}^{\text{ pre}} \cdot V_{c},
\end{split}
\end{equation}
where \( Q_{s}^{\text{pre}} \), \( K_{s}^{\text{pre}} \), and \( V_{s} \) are the query, key, and value features for self-attention, respectively, and \( Q_{c}^{\text{pre}} \), \( K_{c}^{\text{pre}} \), and \( V_{c} \) are those for cross-attention. \( d \) denotes the feature dimension for attention scaling.
The terms \( \textit{s}_{t}^{\text{ pre}} \) and \( \textit{c}_{t}^{\text{ pre}} \) represent the precomputed self- and cross-attention maps at time step \( t \), obtained from DDIM inversion over the preview frames.
% Self-attention has been shown to more reliably capture motion information \cite{Fatezero}, helping preserve both motion integrity and spatial details in the preview frames. Meanwhile, as illustrated in Fig.~\ref{fig:figure1}, the cross-attention maps provide a semantic layout of the preview frames \cite{Prompt_to_prompt}, further reinforcing structural consistency throughout the denoising process. These maps act as guiding signals and, through the proposed IBA, effectively improve both temporal coherence and structural alignment during generation.
Self-attention captures motion and spatial details \cite{Fatezero}, while cross-attention maps provide semantic layouts \cite{Prompt_to_prompt}, reinforcing structure. Through IBA, these signals jointly improve temporal coherence and structural alignment.
\noindent \textbf{Reference-Anchored Self-Attention (RA-SA).}
% The control branch and IBA effectively ensure that the identity's pose in the generated video follows the pose sequence, and the identity's appearance adheres to the reference image. However, since we use image diffusion models \cite{stablediffusion}, the generated video may suffer from temporal inconsistency across frames. To mitigate this issue, we draw inspiration from causal spatial self-attention \cite{Pix2video, Fatezero, Tokenflow} and propose Reference-Anchored Self-Attention (RA-SA) , which enforces consistency among generated frames by anchoring them to the reference image.
The control branch and IBA ensure pose alignment and identity preservation, but image diffusion models \cite{stablediffusion} still suffer from temporal inconsistency. To address this, we propose Reference-Anchored Self-Attention (RA-SA), inspired by causal spatial attention \cite{Pix2video,Fatezero}, which enforces frame consistency by anchoring to the reference image.
Specifically, $\text{S\scalebox{0.75}{ELF}-A\scalebox{0.75}{TTENTION}}(Q, K, V)$ for the latent code \( z^{i}_{t} \) of the \( i \)-th frame \((i \in [1, N])\) at denoising time step \( t \in [1, T] \) is implemented in RA-SA as follows:
\begin{equation}
\begin{split}
Q=W^{Q} z^{i}_{t}, K=W^{K}\left[z^{i}_{t} ; z^{a}_{t}\right], V=W^{V}\left[z^{i}_{t} ; z^{a}_{t}\right],
\end{split}
\label{equ:equation4}
\end{equation}
where $W^Q$, $W^K$, and $W^V$ represent the projection matrices from the fixed U-Net, and $\left[\cdot\right]$ denotes the concatenation operation, $z^{i}_{t}$ and $z^{a}_{t}$ denote the latents of current frame and anchor frame. In Pixel2Video \cite{Pix2video} and FateZero \cite{Fatezero}, $a$ is assigned values of $1$ and $\left\lfloor \frac{N}{2} \right\rfloor$, respectively. In our work, to better align with the goal of HIA, we instead designate ${I_{ref}}$ as anchor frame, serving as a consistent reference for the attention mechanism.
It is worth noting that, during the DDIM inversion of preview frames, we also use RA-SA to replace the original self-attention mechanism. This modification results in a change in the dimensions of the self-attention map \( \textit{s}_{t}^{\text{ pre}} \), from $R^{hw \times hw}$ to $R^{hw \times 2hw}$. Due to the combined use of IBA and RA-SA, during the denoising process, both the query and key features in the self-attention mechanism are taken from the DDIM inversion attention maps of the preview frames, while the value features are computed in real-time using the current latent and reference latent, as shown in Eq.\ref{equ:equation4}.
% Ablation study shows that this design significantly improves frame consistency.
\subsection{Preview Generation Strategy}
\label{sub:PreviewGenerationStrategy}
\begin{figure}[tb] \centering
\includegraphics[width=0.45\textwidth]{figures/Fig3_v2.pdf}
\caption{Overview of the Preview Generation Strategy.}
\vspace{-2mm}
\label{fig:figure3}
\vspace{-3mm}
\end{figure}
% Initial noise serves as the starting point in the denoising process of diffusion models, critically influencing the generated images. Variations in initial noise influence the diversity of outputs and can incorporate additional information to guide the generation process. Many video editing methods \cite{Pix2video, TuneAvideo, zhang2023controlvideo} utilize DDIM inversion to generate initial noise for high-quality video editing. The HIA task can be viewed as video editing, where each frame is derived from the reference image based on the provided pose sequence. Therefore, using the initial noise obtained by inverting preview frames that closely resemble the target video can improve the fidelity of generated images. Motivated by this, we propose the Preview Generation Strategy, a general approach to generate preview frames closely aligned with the target frames, enabling more accurate generation using powerful pretrained models.
Initial noise serves as the starting point in the denoising process of diffusion models. Many video editing methods use DDIM inversion to obtain noise for high-quality editing. Since HIA can be viewed as video editing, where each frame is derived from a reference image and pose, using inverted noise from preview frames similar to the target improves fidelity. Motivated by this, we propose the Preview Generation Strategy, a general approach to generate preview frames closely aligned with the target frames, enabling more accurate generation using powerful pretrained models.
% The preview generation pipeline is illustrated in Figure~\ref{fig:figure3}. Specifically, we use MasaCtrl \cite{Masactrl} with the T2I-Adapter \cite{T2iAdapter} to generate the pose-conditioned image $(a)$. MasaCtrl is a general image editing model capable of real image editing without fine-tuning. T2I-Adapter, similar to ControlNet, is a plug-and-play module that provides additional guidance to pre-trained text-to-image models without altering their original network topology or generation capabilities.
% The initial noise of MasaCtrl with T2I-Adapter is obtained from the reference image ${I_{ref}}$ via DDIM inversion, while the T2I-Adapter's conditioning image comes from the pose sequence $p_i$. In the generated image $(a)$, the identity’s appearance closely resembles ${I_{ref}}$, but the background differs noticeably from the original. To address this, we use the Grounded-SAM model \cite{GroundedSAM} to retain the foreground, producing image $(b)$, referred to as the preview foreground.
% To obtain the clean background from ${I_{ref}}$, we first use the Grounded-SAM model to remove the foreground identity, resulting in a background image with a missing region $(c)$. Next, we apply an image inpainting method, MAT \cite{Mat}, to fill the missing region, generating image $(d)$, referred as the preview background. However, several unrealistic pixel areas appear in the previously missing region of the preview background. We attribute this to a domain gap between the MAT algorithm and our specific application scenario.
% However, the suboptimal regions in the preview background do not degrade our preview frames, as we perform a pixel-wise mix-up of the preview foreground and preview background. The foreground, placed on top, effectively masks the suboptimal areas in the background. The final generated preview frames, referred to as image $(e)$, are thus produced.
The preview generation pipeline is illustrated in Figure~\ref{fig:figure3}. We first use MasaCtrl \cite{Masactrl} with the T2I-Adapter \cite{T2iAdapter} to generate the pose-conditioned image $(a)$.
% MasaCtrl enables real image editing without fine-tuning, while T2I-Adapter provides additional guidance to pretrained text-to-image models without changing their structure or capacity.
The initial noise of this process is obtained from the reference image $I_{ref}$ via DDIM inversion, and the conditioning image comes from the pose sequence $p_i$. In the generated image $(a)$, the appearance matches $I_{ref}$ but the background differs. To address this, Grounded-SAM \cite{GroundedSAM} is used to preserve the foreground, producing image $(b)$, the preview foreground.
To extract only the background, we first apply Grounded-SAM to remove the identity from $I_{ref}$, producing image $(c)$ with a missing region. The missing part is then filled using the MAT inpainting method \cite{Mat}, generating image $(d)$, the preview background. Although some unrealistic pixels remain due to a domain gap between MAT and our scenario, this does not affect the final result. By mixing the preview foreground and background at the pixel level, the foreground masks imperfect background regions. The final preview frames are obtained as image $(e)$.
For more preview frame examples, please refer to our project page.
% Figure~\ref{fig:figure4} shows preview frames across multiple datasets.
% \begin{figure}[h] \centering
% \includegraphics[width=0.45\textwidth]{figures/Fig4.pdf}
% \caption{Visualizations of the preview frames produced by the Preview Generation Strategy.}
% \label{fig:figure4}
% \vspace{-2mm}
% \end{figure}
% The resulting preview frames are then inverted using DDIM inversion to produce the initial noise required for subsequent denoising process. More visual results of the preview background, preview foreground, and preview frames are included in the supplementary material.
\begin{table*}[t] \centering
\newcommand{\Frst}[1]{\textbf{#1}}
\newcommand{\Scnd}[1]{\underline{#1}}
\newcommand{\Trd}[1]{\textcolor{darkgreen}{\textbf{#1}}}
\caption{Quantitative comparison on TikTok and TED-Talks datasets. In the table, $a/b$ denotes results on TikTok and TED-Talks, respectively. \textbf{Bold} text indicates the best result, while \underline{underlined} text indicates the second-best. ↑ indicates that higher values are better. PoseAnimate provides limited experimental details and visual results, with its results (marked in gray) included for reference only. }
\label{tab:table1}
\input{tables/Table_1}
\vspace{-1mm}
\end{table*}
% \begin{table*}[t] \centering
% \newcommand{\Frst}[1]{\textbf{#1}}
% \newcommand{\Scnd}[1]{\underline{#1}}
% \newcommand{\Trd}[1]{\textcolor{darkgreen}{\textbf{#1}}}
% \caption{Quantitative comparison on TED-talks dataset.}
% \label{tab:table2}
% \input{tables/Table_2}
% \vspace{-1mm}
% \end{table*}
\begin{figure*}[tb] \centering
\includegraphics[width=0.92\textwidth]{figures/Fig6_v4.pdf}
\caption{Qualitative comparisons between FreeAnimate and baselines on TikTok (top row) and TED-Talks (bottom row) datasets. The conditioning pose maps are overlaid in the top corner of the generated frames.} \label{fig:figure5}
\vspace{-5mm}
\end{figure*}
\section{Experiments}
\label{sec:Experiments}
\textbf{Baselines and Benchmarks.}
We chose a diverse set of state-of-the-art HIA methods for comparison.
MRAA \cite{MRAA} and TPS \cite{TPS} are state-of-the-art GAN-based methods. Diffusion-based methods are grouped by training data: (1) DisCo \cite{Disco}, MagicPose \cite{MagicPose}, MagicAnimate \cite{Magicanimate}, and AnimateAnyone \cite{Animateanyone} use only public datasets; (2) Champ \cite{Champ}, MimicMotion \cite{Mimicmotion}, and StableAnimator \cite{StableAnimator} combine public and private data; (3) PoseAnimate \cite{PoseAnimate} is training-free, but since no code or TikTok/TED-Talks visual results are available, we report its results in gray for reference only.
% Most quantitative comparisons were based on statistical data from the original literature and FYPv2 \cite{FYPv2}, except StableAnimator, Champ and MimicMotion on TED-talks dataset, which is obtained by re-running the official code.
% All visualization results of AnimateAnyone were obtained using the open-source code provided by Moore Threads \footnote{\href{https://github.com/MooreThreads/Moore-AnimateAnyone}{\textit{https://github.com/MooreThreads/Moore-AnimateAnyone}}} and the DensePose \cite{Densepose} conditioning maps required for MagicAnimate \cite{Magicanimate} are generated using the codebase provided by Flode-Labs \footnote{\href{https://github.com/Flode-Labs/vid2densepose}{\textit{https://github.com/Flode-Labs/vid2densepose}}}.
% \noindent \textbf{Benchmark Dataset.}
% TikTok dataset \cite{TikTok} comprises 350 short-duration (10$\sim$15 seconds) dance videos, mostly featuring a single individual. For this benchmark, we employ the same test set as DisCo \cite{Disco}.
% TED-talks \cite{MRAA} dataset comprises a large collection of video clips featuring primarily the upper body of speakers. We randomly selected 30 video clips from the official test set to conduct our evaluation.
% EverybodyDanceNow dataset \cite{EverybodyDanceNow} consists of full-body videos of five subjects. Using this dataset, we aim to evaluate the generalization of our method to full-body motions captured in unconstrained environments. Further details about datasets and evaluation metrics can be found in the supplementary material.
The TikTok dataset \cite{TikTok} contains 350 short dance videos (10 to 15 seconds), mostly with a single subject. For this benchmark, we use the same test set as DisCo \cite{Disco}. The TED-Talks dataset \cite{MRAA} includes a large collection of speaker videos focusing on the upper body, from which we randomly selected 30 clips for evaluation. The EverybodyDanceNow dataset \cite{EverybodyDanceNow} consists of full-body videos of five subjects and is used to assess the generalization of our method to full-body motions.
\noindent
\textbf{Implementation Details.}
% We use DWPose \cite{DWpose} to extract the pose sequence from the driving video, and adopt the pose alignment algorithm proposed in StableAnimator \cite{StableAnimator} to align the driving pose sequence with the reference pose. For each video, we sample the first frame of the video clip as the reference and all others as driving frames. Both reference and driving frames are center-cropped at the same position along the height dimension with the aspect ratio of 1 before resizing to 512 × 512.
% Due to memory and VRAM constraints, each animation sequence is generated with 8 frames at a resolution of 512 × 512 pixels. To produce longer videos, we followed the long-video animation strategy proposed in MagicAnimate \cite{Magicanimate}.
% For our experiments, we configured the DDIM sampler with the default settings of denoising steps T = 50 and a classifier-free guidance scale of $7.5$. We apply Inversion-Boosted Attention in the interval $t \in[0.5 \times T, T]$ of the denoising step with total timestep $T = 50$, while the Reference-Anchored Self-Attention is applied across the entire denoising step.
% All experiments are conducted on an NVIDIA 3090 GPU, with an approximate time of 5 seconds required to complete 50 denoising steps for a single image.
% Furthermore, xFormers \cite{xFormers2022} are employed to optimize memory efficiency.
We use DWPose \cite{DWpose} for pose extraction and StableAnimator’s \cite{StableAnimator} algorithm for alignment. The first frame is the reference, and others are driving frames, resized to 512 × 512.
The DDIM sampler is configured with T = 50 steps and a guidance scale of $7.5$. IBA is applied from $t \in[0.5 \times T, T]$, and RA-SA throughout. Experiments are run on an NVIDIA 3090 GPU, with xFormers for memory optimization. More details are available on our project page.
% \textcolor{red}{More details are available on our project page.}
\subsection{Qualitative and Quantitative Results}
\label{sub:all_results}
\textbf{Qualitative Comparison.}
In Figure~\ref{fig:figure5}, we present a qualitative comparison between FreeAnimate and other methods.
% Additional visualization results can be found in the supplementary material, where we provide more other domain examples to demonstrate generalization capability.
FreeAnimate demonstrates competitive performance against training-based approaches, with visually superior generation quality compared to the well-established baseline, DisCo \cite{Disco}. Since our method is training-free, it exhibits strong generalization across various datasets. Additionally, preview frames provide a strong prior for the background in generated frames, ensuring high alignment between the background of the generated frames and that of the reference image.
% However, we observe that nearly all HIA methods, including ours, struggle with generating fine details in hands and facial areas, likely due to the inherent randomness in diffusion models.
% \noindent \textbf{Quantitative Comparison.}
% Due to motion blur and low image quality in some samples from the TikTok and TED-talks dataset, the predicted pose maps from DWPose often contain noticeable errors, such as missing body parts or inaccurate keypoint localization—particularly in the hands. These flawed pose maps can significantly degrade the generation quality. Therefore, during evaluation, we excluded samples with pose maps that were subjectively assessed as low quality.
% Table~\ref{tab:table1} presents the quantitative comparison results of FreeAnimate on the TikTok dataset \cite{TikTok}.
% Our method outperforms the training-based baseline DisCo \cite{Disco} on all metrics, with significant improvements in SSIM (+9.4\%), PSNR (+72.3\%), and FVD (+31.4\%). Moreover, the visual quality of our generated images is comparable to models trained with large-scale datasets. For instance, our method surpasses MagicAnimate \cite{Magicanimate} in FID, SSIM, and L1 error, and achieves better overall performance than MimicMotion \cite{Mimicmotion} across most metrics.
% However, there remains a performance gap in FID-VID and FVD compared to the training based methods like Champ \cite{Champ}. We attribute this discrepancy to the fact that our method does not rely on video data for training, which inevitably impacts the temporal consistency of the generated results.
% Table~\ref{tab:table1} also presents the quantitative comparison results of FreeAnimate on the TED-talks dataset \cite{MRAA}.
% Since MRAA \cite{MRAA}, TPS \cite{TPS}, and AnimateAnyone \cite{Animateanyone} are trained on the TED-Talks dataset, their performance demonstrates a clear advantage. In contrast, other methods exhibit a noticeable drop in most metrics on TED-Talks compared to TikTok, as the TED-Talks data is not included in their training process. Being training-free, FreeAnimate is less affected by data distribution differences, allowing it to maintain consistent performance across both datasets.
% We attribute the performance degradation of most methods to domain shifts between datasets. Specifically, TED-Talks images often feature larger foreground subjects, requiring stronger modeling of fine-grained human details, such as facial structures and distinguishable fingers, which is one of the key challenges in HIA.
\noindent \textbf{Quantitative Comparison.}
To reduce bias from erroneous DWPose estimates (e.g., missing hands), we exclude samples with evidently low-quality pose maps. As summarized in Table~\ref{tab:table1}, FreeAnimate is highly competitive despite being training-free. On TikTok dataset, it ranks 2nd or 3rd on FID/SSIM/L1/FVD, outperforming DisCo and MagicAnimate on most image metrics.
Table~\ref{tab:table1} also presents comparisons on the TED-Talks dataset.
Since MRAA \cite{MRAA}, TPS \cite{TPS}, and AnimateAnyone \cite{Animateanyone} are trained on the TED-Talks dataset, their performance demonstrates a clear advantage. In contrast, other methods exhibit a noticeable drop on TED-Talks.
%Being training-free, FreeAnimate is less affected by data distribution differences, allowing it to maintain consistent performance across both datasets.
% Since we utilize fundamental and widely applicable models (SD v1.5 and ControlNet v1.1), our method does not achieve superior performance in terms of metrics. However, the proposed framework is training-free and features loosely coupled modules, allowing future improvements by replacing existing models with more advanced ones (e.g., SDXL) for enhanced generation results.
%Since our method builds on standard models (SD v1.5, ControlNet v1.1), it does not achieve state-of-the-art metrics. However, it is training-free and modular, enabling future gains by replacing components with stronger models.
Being training-free, FreeAnimate is less affected by data distribution shifts, maintaining consistent performance across datasets. Built on standard models (SD v1.5, ControlNet v1.1), it does not reach state-of-the-art metrics, but its training-free and modular design enables future gains by replacing components with stronger models.
\begin{comment}
\subsection{Ablation Study}
All ablation studies are conducted on TikTok dataset.
\renewcommand{\arraystretch}{1.1} % 设置行距为默认的1.5倍
\begin{table}[tb]\centering
\caption{Ablation study of the Preview Generation Strategy.}
\label{tab:table3}
\resizebox{0.4\textwidth}{!}{
\small
\begin{tabular}{l*{9}{c}}
\toprule
Method & FID↓ & SSIM↑ & FVD↓ \\
\midrule
FreeAnimate \textit{w/o.} preview & 60.07 & 0.349 & 320.91 \\
FreeAnimate \textit{w/.} driving & 23.54 & 0.786 & 157.55 \\
\rowcolor{gray!15}
FreeAnimate (Ours) & 27.82 & 0.731 & 201.18 \\ \hline
MagicPose \cite{MagicPose} & 25.50 & 0.752 & 216.01 \\
MagicPose \textit{w/.} preview & 24.61 & 0.759 & 180.49 \\
\bottomrule
\end{tabular}
}
\vspace{-2mm}
\end{table}
\renewcommand{\arraystretch}{1.1} % 设置行距为默认的1.5倍
\begin{table}[tb]\centering
\caption{Ablation study of architecture design.}
\label{tab:table4}
\resizebox{0.45\textwidth}{!}{
\normalsize
\begin{tabular}{l*{9}{c}}
\toprule
Method & FID↓ & SSIM↑ & Face-Sim↑ & FVD↓ \\
\midrule
\textit{w/o.} Control Branch & 38.92 & 0.637 & 0.514 & 237.89 \\
\textit{w/o.} IBA & 44.34 & 0.564 & 0.584 & 256.10 \\
\textit{w/o.} RA-SA & 51.20 & 0.442 & 0.406 & 302.02 \\
\rowcolor{gray!15}
FreeAnimate (Ours) & 27.82 & 0.731 & 0.654 & 201.18 \\
\bottomrule
\end{tabular}
}
\vspace{-3mm}
\end{table}
\noindent
\textbf{Preview Generation Strategy.}
The preview frames generated by the Preview Generation Strategy contribute to the initial noise, and the attention maps produced during the DDIM inversion process are utilized at most steps of the denoising process. Without the Preview Generation Strategy, the model’s input reverts to a combination of randomly initialized noise. In this setting, Inversion-Boosted Attention is no longer enabled, leaving only Reference-Anchored Self-Attention to enhance temporal consistency. As shown in Table~\ref{tab:table3} (Row 1), this configuration results in a noticeable degradation in generation quality.
Replacing the preview frames with frames from the driving video leads to a marked improvement in generation quality, as shown in Table~\ref{tab:table3} (Row 2). This result suggests a positive correlation between the quality of the preview frames and the overall generation quality.
Additionally, when we applied the inverted noise generated by the Preview Generation Strategy as the initial noise for MagicPose \cite{MagicPose}, the results, presented in Table~\ref{tab:table3} (Row 5) as "MagicPose \textit{w/.} preview," show consistent performance improvements over the original MagicPose. This further highlights the effectiveness of the Preview Generation Strategy, suggesting that it can provide a strong starting point for other training-based methods.
\noindent
\textbf{Architecture Design.}
As introduced in Section \ref{sub:NetworkArchitecture}, the model's architecture design consists of three components: Control Branch, Inversion-Boosted Attention (IBA), and Reference-Anchored Self-Attention (RA-SA).
To evaluate the contribution of different modules to identity preservation, we introduce a new metric, \textbf{Face-Sim}. Specifically, we extract face embeddings from the generated images under different settings using ArcFace \cite{Arcface}, and compute the average cosine similarity between these embeddings and that of the reference image. The resulting similarity score is reported as the Face-Sim.
When the Control Branch is removed, the image quality metrics (FID, SSIM and Face-Sim) show a noticeable decline, as seen in Table~\ref{tab:table4} (Row 1). In contrast, video-related metrics (FID-VID and FVD) experience only a relatively slight decrease. The generated results reveal a reduced ability to control fine details, such as facial expressions and hand positioning, though the overall motion remains aligned with the pose sequence. We attribute this to the inverted initial noise from the preview frames, which provides prior information about the identity’s pose, as well as to the IBA, which injects pose details during the denoising process.
As shown in Table~\ref{tab:table4} (Row 2), removing IBA leads to a clear decline in frame quality and temporal consistency, as this module helps the U-Net maintain layout consistency with the preview frames and provides a strong prior for background stability across frames.
RA-SA provides temporal consistency for the generated video and incorporates explicit priors from the reference image. In its absence, FreeAnimate experiences a significant drop in generation quality and identity preservation, as shown in Table~\ref{tab:table4} (Row 3).
\end{comment}
\subsection{Ablation Study}
\begin{table}[tb]\centering
\caption{Ablation study of FreeAnimate.}
\label{tab:table_ablations}
\resizebox{0.4\textwidth}{!}{
\small
\begin{tabular}{l | l*{5}{c}}
\toprule
Method & FID↓ & SSIM↑ & FVD↓ \\
\midrule
FreeAnimate \textit{w/o.} preview & 50.07 & 0.549 & 260.91 \\
FreeAnimate \textit{w/.} driving & 23.54 & 0.817 & 147.55 \\
\rowcolor{gray!15} \textbf{FreeAnimate (Ours)} & 27.82 & 0.781 & 170.18 \\
MagicPose \cite{MagicPose} & 25.50 & 0.752 & 216.01 \\
MagicPose \textit{w/.} preview & 24.61 & 0.760 & 180.49 \\
\midrule
\textit{w/o.} Control Branch & 35.92 & 0.671 & 190.89 \\
\textit{w/o.} IBA & 39.34 & 0.634 & 200.10 \\
\textit{w/o.} RA-SA & 39.20 & 0.602 & 259.02 \\
\bottomrule
\end{tabular}
}
\vspace{-5mm}
\end{table}
All ablation studies are conducted on the TikTok dataset.
The Preview Generation Strategy plays a crucial role in generating high-quality videos by providing initial noise from preview frames, aiding the denoising process with attention maps. Removing this strategy, as shown in Row 1 of Table~\ref{tab:table_ablations}, leads to degraded generation quality due to the use of randomly initialized noise. Using frames from the driving video improves generation significantly (Row 2). This suggests that high-quality preview frames positively impact overall generation. Additionally, applying this strategy to MagicPose results in improved performance compared to the original (Row 5).
Regarding architecture design, FreeAnimate's performance relies on three key components: Control Branch, IBA, and RA-SA. Removing the Control Branch (Row 6) causes a noticeable drop in image quality metrics, though video consistency remains intact. The IBA (Row 7) stabilizes the layout and ensures background consistency, while RA-SA (Row 8) preserves temporal consistency and identity. Removing any of these modules leads to significant performance degradation in both identity preservation and temporal coherence.
% Overall, the effectiveness of the Preview Generation Strategy and each architectural component is clear, with each contributing significantly to video generation quality.
% \subsection{Computational cost.}
FreeAnimate trades higher inference time for zero training cost. We report a stage-wise latency/peak-VRAM breakdown under a standardized setting in Table~\ref{tab:efficiency}. The pipeline is modular: preview/segmentation/inpainting can be replaced by faster alternatives, and diffusion can be accelerated via fewer-step sampling and standard engineering optimizations.
\begin{table}[t]
\centering
\caption{Computational cost analysis.}
\label{tab:efficiency}
\resizebox{0.4\textwidth}{!}{
\begin{tabular}{l l | c c}
\hline
Stage & Key modules & Time/frame & Peak VRAM \\
\hline
1 & Preview generation & 3103 ms & 4623 MB \\
2 & DDIM inversion & 1043 ms & 3017 MB \\
3 & Denoising & 1350 ms & 5561 MB \\
\hline
& All above & 5496 ms & 5561 MB \\
\hline
\end{tabular}
}
\vspace{-5mm}
\end{table}
\section{Conclusion}
\label{sec:Conclusion}
We present FreeAnimate, a training-free framework for HIA, addressing limitations associated with data- and resource-intensive HIA methods. By integrating a Preview Generation Strategy with a training-free model architecture, FreeAnimate achieves high fidelity in pose-guided human image animation without relying on extensive training datasets. Through comprehensive evaluations, FreeAnimate consistently outperforms training-free competitors and exhibits generalization across various datasets, illustrating its practical value in diverse applications.
Future work may focus on enhancing fine detail generation, particularly in facial expressions and hands, and exploring HIA under dynamic and complex backgrounds.
\clearpage
\section{Acknowledgements}
This work is supported by the National Natural Science Foundation of China (U23B2030).
% References should be produced using the bibtex program from suitable
% BiBTeX files (here: strings, refs, manuals). The IEEEbib.bst bibliography
% style file from IEEE produces unsorted bibliography list.
% -------------------------------------------------------------------------
\bibliographystyle{IEEEbib}
\bibliography{refs}
\end{document}
% \renewcommand{\arraystretch}{1.1} % 设置行距为默认的1.5倍
% \resizebox{0.9\textwidth}{!}{
% \scriptsize
% \begin{tabular}{l||l||l||*{5}{c}|*{2}{c}}
% \toprule
% & Training Data & Method & \textbf{FID↓} & \textbf{SSIM↑} & \textbf{PSNR↑} & \textbf{LPIPS↓} & \textbf{L1↓} & \textbf{FID-VID↓} & \textbf{FVD↓} \\ \hline
% % \multirow{4}{*}{GAN} & \multirow{4}{*}{Public} & MRAA* \cite{MRAA} & \textcolor{gray!60}{54.47} & \textcolor{gray!60}{0.672} & \textcolor{gray!60}{29.39} & \textcolor{gray!60}{0.296} & \textcolor{gray!60}{3.21E-04} & \textcolor{gray!60}{66.36} & \textcolor{gray!60}{284.82} \\
% % & & TPS* \cite{TPS} & \textcolor{gray!60}{53.78} & \textcolor{gray!60}{0.673} & \textcolor{gray!60}{29.18} & \textcolor{gray!60}{0.299} & \textcolor{gray!60}{3.23E-04} & \textcolor{gray!60}{72.55} & \textcolor{gray!60}{306.17} \\
% \multirow{2}{*}{GAN} & \multirow{2}{*}{Public} & MRAA \cite{MRAA} & 85.49 & 0.646 & 28.39 & 0.337 & 4.61E-04 & 71.97 & 468.66 \\
% & & TPS \cite{TPS} & 140.37 & 0.560 & 28.17 & 0.449 & 6.17E-04 & 142.52 & 800.77 \\ \midrule
% \multirow{8}{*}{SD} & \multirow{4}{*}{Public} & DisCo \cite{Disco} & \Trd{30.75} & 0.668 & 16.55 & 0.292 & 3.78E-04 & 59.90 & 292.80 \\
% & & MagicPose \cite{MagicPose} & \Frst{25.50} & \Trd{0.752} & \Trd{29.53} & 0.292 & \Frst{0.81E-04} & 46.30 & 216.01 \\
% & & MagicAnimate \cite{Magicanimate} & 32.09 & 0.714 & 29.16 & \Trd{0.239} & 3.13E-04 & \Scnd{21.75} & 179.07 \\
% & & AnimateAnyone \cite{Animateanyone} & - & 0.718 & \Scnd{29.56} & 0.285 & - & - & \Trd{171.90} \\ \cline{2-10}
% & \multirow{3}{*}{\begin{tabular}[c]{@{}c@{}}Public+Private\end{tabular}} & Champ \cite{Champ} & - & \Frst{0.802} & \Frst{29.91} & \Scnd{0.234} & 2.94E-04 & \Frst{21.07} & \Scnd{160.82} \\
% % & & MimicMotion \cite{Mimicmotion} & \Trd{28.03} & 0.713 & 25.87 & 0.314 & \Scnd{1.02E-04} & 43.56 & 410.55 \\ \cline{2-10}
% & & MimicMotion \cite{Mimicmotion} & - & 0.601 & 14.44 & 0.416 & 5.85E-04 & - & 326.57 \\
% & & StableAnimator \cite{StableAnimator} & - & \Scnd{0.801} & 20.66 & \Frst{0.232} & \Trd{2.87E-04} & - & \Frst{140.62} \\ \cline{2-10}
% & \multirow{2}{*}{None} & PoseAnimate \cite{PoseAnimate} & \textcolor{gray!60}{31.47} & - & - & - & \textcolor{gray!60}{3.06E-04} & \textcolor{gray!60}{63.26} & \textcolor{gray!60}{286.33} \\
% % & & FreeAine (Ours) & 29.01 & 0.711 & 23.52 & 0.305 & 2.53E-04 & 52.27 & 201.18 \\
% & & FreeAnimate (Ours) & \Scnd{27.82} & 0.731 & 28.52 & 0.289 & \Scnd{2.23E-04} & \Trd{42.27} & 201.18 \\
% % & & FreeAine (Ours) & 29.01 & 0.731 & 26.52 & 0.305 & 2.23E-04 & 42.27 & 201.18 \\
% \bottomrule
% \end{tabular}
% }
\renewcommand{\arraystretch}{1}
\resizebox{0.9\textwidth}{!}{
\tiny
% \scriptsize
% \small
% \normalsize
\setlength{\tabcolsep}{3pt}
\begin{tabular}{l||l||l||*{5}{c}|*{1}{c}}
\toprule
& Training Data & \textbf{Method} & \textbf{FID↓} & \textbf{SSIM↑} & \textbf{PSNR↑} & \textbf{LPIPS↓} & \textbf{L1↓ (E-04)} & \textbf{FVD↓} \\
\hline
\multirow{2}{*}{GAN} & \multirow{2}{*}{Public}
& MRAA \cite{MRAA} & 85.49 / 50.36 & 0.646 / 0.762 & 28.39 / 31.90 & 0.337 / 0.266 & 4.61 / \Scnd{0.50} & 468.66 / 493.02 \\
& & TPS \cite{TPS} & 140.37 / 23.71 & 0.560 / 0.771 & 28.17 / \Scnd{32.30} & 0.449 / 0.252 & 6.17 / \Frst{0.49} & 800.77 / 260.67 \\
\hline
\multirow{8}{*}{SD} & \multirow{4}{*}{Public}
& DisCo \cite{Disco} & 30.75 / 75.48 & 0.668 / 0.575 & 16.55 / 27.99 & 0.292 / 0.309 & 3.78 / 1.2 & 292.80 / 393.04 \\
& & MagicPose \cite{MagicPose} & \Frst{25.50} / \Scnd{23.39} & 0.752 / 0.723 & 29.53 / 30.08 & 0.292 / 0.236 & \Frst{0.81} / 0.81 & 216.01 / 214.23 \\
& & MagicAnimate \cite{Magicanimate} & 32.09 / 41.58 & 0.714 / 0.529 & 29.16 / 28.28 & 0.239 / 0.310 & 3.13 / 1.73 & 179.07 / 223.54 \\
& & AnimateAnyone \cite{Animateanyone} & - / 25.93 & 0.718 / \Frst{0.832} & \Scnd{29.56} / \Frst{33.91} & 0.285 / \Frst{0.159} & - / - & 171.90 / \Frst{80.50} \\
\cline{2-9}
& \multirow{3}{*}{Public+Private}
& Champ \cite{Champ} & - / 64.96 & \Frst{0.802} / 0.625 & \Frst{29.91} / 21.35 & \Scnd{0.234} / 0.307 & 2.94 / 1.56 & \Scnd{160.82} / 267.03 \\
& & MimicMotion \cite{Mimicmotion} & - / 76.93 & 0.601 / 0.580 & 14.44 / 19.77 & 0.416 / 0.369 & 5.85 / 3.84 & 326.57 / 305.03 \\
& & StableAnimator \cite{StableAnimator} & - / \Frst{22.72} & \Scnd{0.801} / 0.769 & 20.66 / 27.92 & \Frst{0.232} / \Scnd{0.218} & 2.87 / 1.42 & \Frst{140.62} / \Scnd{165.75} \\
\cline{2-9}
& \multirow{2}{*}{None}
& PoseAnimate \cite{PoseAnimate} & \textcolor{gray!60}{31.47 / 21.24} & \textcolor{gray!60}{- / -} & \textcolor{gray!60}{- / -} & \textcolor{gray!60}{- / -} & \textcolor{gray!60}{3.06 / 1.98} & \textcolor{gray!60}{286.33 / 168.02} \\
& & \textbf{FreeAnimate (Ours)} & \Scnd{27.82} / 24.31 & 0.781 / \Scnd{0.803} & 29.26 / 29.18 & 0.244 / 0.231 & \Scnd{2.23} / 1.30 & 170.18 / 169.92 \\
\bottomrule
\end{tabular}
}