DreamActor-M2: Universal Character Image Animation via Spatiotemporal In-Context Learning

Mingshuang Luo123,*, Shuang Liang1,*, Zhengkun Rong1,*, Yuxuan Luo1,†, Tianshu Hu1,§, Ruibing Hou2,§, Hong Chang23, Yong Li4, Yuan Zhang1, Mingyuan Gao1
1Bytedance Intelligent Creation, 2Key Lab of Intell. Info. Process., ICT, CAS,
3University of Chinese Academy of Sciences, 4Southeast University
*Equal Contribution †Project Lead §Corresponding Authors
Paper 🤗 Hugging Face

We introduce DreamActor-M2, a universal character image animation framework that reformulates motion conditioning as a spatiotemporal in-context learning task. Our design harnesses the inherent generative priors of video foundation models while facilitating a critical evolution toward pose-free, end-to-end motion transfer directly from raw videos. This paradigm eliminates the need for explicit pose estimation, enabling DreamActor-M2 to achieve exceptional generalization and high-fidelity results across diverse and complex scenarios.

Abstract

Character image animation aims to synthesize high-fidelity videos by transferring motion from a driving sequence to a static reference image. Despite recent advancements, existing methods suffer from two fundamental challenges: (1) suboptimal motion injection strategies that lead to a trade-off between identity preservation and motion consistency, manifesting as a "see-saw", and (2) an over-reliance on explicit pose priors (e.g., skeletons), which inadequantely capture intricate dynamics and hinder generalization to arbitrary, non-humanoid characters. To address these challenges, we present DreamActor-M2, a universal animation framework that reimagines motion conditioning as an in-context learning problem. Our approach follows a two-stage paradigm. First, we bridge the input modality gap by fusing reference appearance and motion cues into a unified latent space, enabling the model to jointly reason about spatial identity and temporal dynamics by leveraging the generative prior of foundational models. Second, we introduce a self-bootstrapped data synthesis pipeline that curates pseudo cross-identity training pairs, facilitating a seamless transition from pose-dependent control to direct, end-to-end RGB-driven animation. This strategy significantly enhances generalization across diverse characters and motion scenarios. To facilitate comprehensive evaluation, we further introduce AWBench, a versatile benchmark encompassing a wide spectrum of characters categories and motion types. Extensive experiments demonstrate that DreamActor-M2 achieves state-of-the-art performance, delivering superior visual fidelity and robust cross-domain generalization.

Method Overview

Interpolate start reference image.

The schematic overview of proposed DreamActor-M2.

Demo Gallary

Character Images: DreamActor-M2 achieves exceptional versatility by animating a broad spectrum of subjects, ranging from realistic humans to stylized cartoons, and even diverse animal species.

Reference Videos: DreamActor-M2 demonstrates robust end-to-end generalization to non-human motion signals, enabling seamless motion transfers without relying on any intermediate pose priors.

Multiple Characters: DreamActor-M2 effectively synchronizes motion across multiple distinct characters or maps complex multi-person dynamics to new groups while maintaining structural integrity.

Human-Object Interactions: Our end-to-end paradigm enables DreamActor-M2 to master intricate human-object interactions, preserving critical spatial and semantic details that are typically lost in pose-estimated representations.

Cross-Shot Animations: By fully harnessing the intrinsic generative priors of foundation models, DreamActor-M2 excels in cross-morphology mapping, adaptively synthesizing missing motions or pruning redundant signals to align with the target shot requirements.

Complex Motions: By extracting motion cues directly from raw RGB videos, DreamActor-M2 accurately captures fine-grained and highly dynamic motion patterns.

Video Comparison

Our method generates results with fine-grained motions, identity preservation, temporal consistency and high fidelity.

BibTeX

@article{luo2025dreamactor,
  title={DreamActor-M1: Holistic, Expressive and Robust Human Image Animation with Hybrid Guidance},
  author={Luo, Yuxuan and Rong, Zhengkun and Wang, Lizhen and Zhang, Longhao and Hu, Tianshu and Zhu, Yongming},
  journal={arXiv preprint arXiv:2504.01724},
  year={2025}
}