WeLike2Party! In-Context Motion Transfer for Multi-Human Image Animation

Sangeyl Lee; Seunghyun Shin; Seungho Park; Wooseok Jeon; Hae-Gon Jeon

cs.CVSep 29, 2026

WeLike2Party! In-Context Motion Transfer for Multi-Human Image Animation

Authors: Sangeyl Lee, Seunghyun Shin, Seungho Park, Wooseok Jeon, Hae-Gon Jeon

Organizations: Department of Artificial Intelligence, Yonsei University · AI Graduate School, GIST

Abstract

Human image animation aims to transfer motion from a driving video to subjects in a reference image. Despite remarkable progress in video generation, achieving high-fidelity animation of multiple interacting subjects remains a challenge. Many existing approaches rely on explicit motion representations such as 2D skeletons or parametric body meshes and struggle to preserve identity-motion binding under inter-person occlusion. To address this limitation, we propose WeLike2Party, a multi-human animation framework built on direct in-context video conditioning without explicit pose or mesh extraction at inference. We further introduce Reference Asymmetric RoPE Conditioning to preserve fine-grained appearance details, and Identity Binding Supervision to associate each reference identity with its intended motion trajectory. To support cross-identity training, we construct MotionTwin, a large-scale synthetic dataset comprising 14.4K cross-identity video pairs with shared subject and camera motions, totaling 84.3 hours of photorealistic video. We additionally present MotionTwin-Bench, a cross-identity benchmark specifically designed to evaluate subject-level visual fidelity and identity-motion binding. Extensive experiments on MotionTwin-Bench and real-world videos demonstrate that WeLike2Party outperforms recent state-of-the-art methods in subject-level visual fidelity, identity-motion binding, and overall perceptual quality, particularly in multi-person interactions with substantial occlusion.

Figures & tables

Figure 1: Human image animation with WeLike2Party . Given a reference image and a driving video, WeLike2Party animates the reference subjects while preserving their appearance and identity-motion correspondence. Our framework enables the creation of both solo and group dance videos.

Figure 2: MotionTwin curation pipeline. We collect and filter diverse SMPL-X motions, sample different avatar appearances, and render cross-identity video pairs with shared motion. Multi-person pairs additionally provide reference images and instance maps for identity correspondence.

Figure 3: Overview of WL2P. WL2P jointly conditions a DiT on the reference image, driving video, and noisy target video. RARC preserves fine-grained reference details through high-resolution encoding, and IBS enforces subject-level identity-motion correspondence via attention supervision.

Method	Full-frame Fidelity				Subject Fidelity			Identity Binding
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FVD $\downarrow$	mPSNR $\uparrow$	mSSIM $\uparrow$	mLPIPS $\downarrow$	IAA $\uparrow$	IAA ${}_{\text{cross}}$ $\uparrow$
MultiAnimate ( Hu et al., 2026b )	20.22	0.7112	0.3283	226.61	18.41	0.7326	0.2029	0.8360	0.8192
Wan-Animate 2 ( Wang et al., 2026 )	17.18	0.5957	0.4609	277.51	15.91	0.6786	0.2969	0.6035	0.5955
SCAIL ( Yan et al., 2026b )	17.42	0.5681	0.3856	164.33	18.11	0.7192	0.2188	0.8514	0.8422
SCAIL-2 ( Yan et al., 2026a )	17.68	0.5775	0.3659	120.75	17.55	0.7097	0.2334	0.8406	0.8335
WeLike2Party (Ours)	25.14	0.7939	0.1123	35.51	25.38	0.8685	0.0691	0.9569	0.9496

Table 1: Quantitative evaluation compared to state-of-the-art methods on MotionTwin-Bench.

Method	VBench		VBench++	User Study
Method	V-Quality $\uparrow$	F-Quality $\uparrow$	I2V-Quality $\uparrow$	Id-Motion Bind $\uparrow$	Id Preserve $\uparrow$	Visual Quality $\uparrow$
MultiAnimate ( Hu et al., 2026b )	79.99	81.40	85.62	2.575 $\pm$ 1.324	2.355 $\pm$ 1.435	2.480 $\pm$ 1.317
Wan-Animate 2 ( Wang et al., 2026 )	81.58	83.09	87.61	2.385 $\pm$ 1.188	2.558 $\pm$ 1.086	2.430 $\pm$ 1.174
SCAIL ( Yan et al., 2026b )	82.84	83.61	85.91	3.153 $\pm$ 1.473	3.338 $\pm$ 1.408	3.205 $\pm$ 1.431
SCAIL-2 ( Yan et al., 2026a )	82.70	84.13	88.41	2.845 $\pm$ 1.212	2.728 $\pm$ 1.224	2.860 $\pm$ 1.253
WeLike2Party (Ours)	83.26	84.63	88.73	4.043 $\pm$ 1.235	4.022 $\pm$ 1.206	4.025 $\pm$ 1.258

Table 2: Quantitative evaluation compared to state-of-the-art methods on the real-world scenario. We report mean and standard deviation for the user study.

Figure 4: Qualitative comparison. MultiAnimate transfers the driving subjects’ hairstyles and faces (a to d), Wan-Animate 2 merges or drops subjects (b to d), SCAIL swaps identities (c, d), and SCAIL-2 copies the driving background (b), alters clothing colors (a, c), and duplicates subjects (c). WL2P preserves each reference subject’s appearance and binds it to the intended motion.

Method	Full-frame Fidelity				Subject Fidelity			Identity Binding
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FVD $\downarrow$	mPSNR $\uparrow$	mSSIM $\uparrow$	mLPIPS $\downarrow$	IAA $\uparrow$	IAA ${}_{\text{cross}}$ $\uparrow$
Base	22.47	0.7022	0.1542	48.81	22.92	0.8122	0.1232	0.8982	0.8891
Base + RARC	23.61	0.7090	0.1264	44.07	24.21	0.8487	0.0851	0.9241	0.9171
Base + HR ref.	22.50	0.6655	0.1395	46.21	23.55	0.8396	0.0947	0.8969	0.8900
Base + IBS	24.49	0.7419	0.1156	36.55	24.70	0.8581	0.0772	0.9303	0.9210
Full (Ours)	25.14	0.7939	0.1123	35.51	25.38	0.8685	0.0691	0.9569	0.9496

Table 3: Ablation study on RARC and IBS. “Base + HR ref.” uses high-resolution reference images with standard RoPE and no coordinate rescaling.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: MotionTwin statistics. Left: hours of rendered video by group size (inner ring) and motion source (outer ring). Right: clips and hours per source.

Figure 6: Overview of MotionTwin. Samples with 1–7 people across diverse scenes.

Method	BindJudge				Vbench		Vbench++
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FVD $\downarrow$	V-Quality $\uparrow$	F-Quality $\uparrow$	I2V-Quality $\uparrow$
MultiAnimate ( Hu et al., 2026b )	20.126	0.5377	0.2283	317.51	80.32	81.64	85.61
Wan-Animate 2 ( Wang et al., 2026 )	20.926	0.5846	0.2447	216.85	81.84	83.11	86.91
SCAIL ( Yan et al., 2026b )	18.389	0.4571	0.2654	263.70	80.97	82.31	86.33
SCAIL-2 ( Yan et al., 2026a )	18.754	0.4608	0.2877	273.18	82.49	84.10	88.91
Wan2.2-Animate-14B ( Cheng et al., 2025 )	17.691	0.4621	0.3353	267.81	81.25	82.53	86.39

Appendix

Table 4: Quantitative evaluation compared to state-of-the-art methods on single-person animation. MotionTwin-Bench metrics are computed on the 75 single-subject windows, and VBench metrics on the 36 single-subject real-world clips.

Figure 7: Qualitative ablation of RARC. Both variants receive the same high-resolution reference image and driving video and differ only in how reference tokens are positioned. With RARC, fingers are rendered distinctly and the face stays intact under fast motion (top, green), whereas assigning native integer RoPE coordinates (Base+HR ref.) yields merged fingers (top, red). The second example (bottom) shows the same trend for hand gestures near the face, where Base+HR ref. blurs the fingers and exhibits identity drift while RARC preserves both.

Figure 8: Additional qualitative results.

Figure 9: Additional qualitative results.

Figure 10: Additional qualitative results.

Explore similar work

Vera: Identity-Faithful Human Subject-to-Video Generation
Jul 22, 2026Yulong Xu, Xinyue Liu, Shujuan Li +6Video Dataset
Kuaishou Technology · Institute of Automation, Chinese Academy of Sciences · Tsinghua University
MultiAnimate: A Unified Framework for Controllable Multi-Character Animation
Jul 15, 2026Zhongyi Zhang, Guangyuan Wang, Li Hu +6Category-Level Object Pose Estimation Unified Framework
School of Cyber Science and Technology, University of Science and Technology of China, Fuxing Road 100, Hefei, 230026, Anhui, China. · Tongyi Lab, Alibaba Group, Dengcai Road 1008, Hangzhou, 310023, Zhejiang, China. · College of Computing and Data Science, Nanyang Technological University, Jurong West, Singapore, 639798, Singapore, Singapore.
Beyond Skeletons: Learning Animation Directly from Driving Videos with Same2X Training Strategy
Jun 5, 2026Yuan Zeng, Yujia Shi, Yuhao Yang +4Beyond Retrieval
1Tsinghua University · 2Harbin Institute of Technology · 3Pengcheng Laboratory

Jul 22, 2026, cs.CVJ/K move · Enter open · S save

Vera: Identity-Faithful Human Subject-to-Video Generation

Yulong Xu, Xinyue Liu, Shujuan Li, huafeng shi+5

Kuaishou Technology · Institute of Automation, Chinese Academy of Sciences · Tsinghua University

Video Dataset

Open paper AskPDF arXiv

Subject-to-video (S2V) generation has made substantial progress in preserving reference subjects across diverse categories, yet generic subject consistency remains insufficient for human-centric generation. A video may appear globally consistent while identity-critical human details still drift across frames, poses, and interactions. This issue becomes more severe in multi-person scenarios, where incorrect identity-role binding leads to subject confusion, attribute swapping, and excessive copying of reference-specific appearance cues. We propose Vera, a unified human-centric S2V framework for single- and multi-person generation. We first construct a million-pair identity-aligned human image-video dataset through person-level cross-clip retrieval, providing explicit identity correspondence and diverse references. Built on this dataset, Vera introduces two complementary designs. Identity-Focal Masked Supervision (IFMS) strengthens identity-aware learning with spatially focused supervision while reducing interference from irrelevant artifacts. Reference-Aware Layer-wise Attention (RALA) regulates how video tokens interact with reference identity cues in the DiT backbone, preserving stable identity anchors and enhancing layer-aware identity readout. Extensive experiments demonstrate that Vera improves human identity consistency, multi-person subject binding, and motion naturalness, while reducing identity confusion and excessive reference-image copying.

Abstract

Explore similar work

Vera: Identity-Faithful Human Subject-to-Video Generation

MultiAnimate: A Unified Framework for Controllable Multi-Character Animation

Beyond Skeletons: Learning Animation Directly from Driving Videos with Same2X Training Strategy