Abstract
Animate Anyone 2 introduces "environment affordance" to character animation, enabling AI-generated characters to interact realistically with their surroundings. By utilizing shape-agnostic masking, a dedicated object guider, and depth-aware pose modulation, the framework achieves state-of-the-art integration between animated subjects and complex backgrounds, overcoming common issues like human-object interaction clipping and spatial inconsistency.
Introduction
While recent diffusion-based models excel at maintaining character appearance consistency during motion, they often treat characters and backgrounds as separate layers, resulting in a "sticker-on-video" effect. Animate Anyone 2 addresses this lack of environment affordance, focusing on three primary challenges: Human-Object Interaction (HOI), spatial consistency in complex movements, and preventing shape leakage from the driving video onto the target character.
Methodology
The framework builds on the original Animate Anyone architecture (ReferenceNet + Denoising 3D UNet) but introduces three major innovations to enhance environmental interaction.
- Environment Formulation & Shape-Agnostic Masks: The environment is provided as a conditional input. To prevent the model from simply memorizing the source person's silhouette, "shape-agnostic masks" create loose boundaries, forcing the model to learn the relationship between the character and the scene rather than just filling a specific shape.
- Object Injection via Spatial Blending: For scenes involving specific objects (like a basketball or a bicycle), a lightweight Object Guider extracts multi-scale features. These are injected into the denoising process via Spatial Blending to ensure interaction dynamics are preserved.
- Depth-wise Pose Modulation: Standard 2D skeletons are augmented with structured depth maps, providing a nuanced understanding of limb-to-limb spatial relationships without the risks of shape leakage associated with raw depth maps.
Datasets and Evaluation
The researchers utilized a massive new internal collection of 100,000 character videos featuring diverse scene types and complex interactions. Benchmarking was performed on the standard TikTok dataset and a real-world test set of 100 videos to evaluate generalizability.
Key Results
Animate Anyone 2 sets a new state-of-the-art (SOTA) across multiple metrics. On the TikTok benchmark, it achieved an FVD (Frechet Video Distance) of 144.65, significantly lower than the original Animate Anyone (171.90).
Qualitative evaluations show superior performance in scenarios involving complex physics, such as riding horses, playing instruments, or using gym equipment. In these cases, previous models typically showed significant "sliding" or limb distortion, whereas Animate Anyone 2 maintains structural integrity and environmental grounding.
Conclusion and Future Work
By shifting from character-centric to interaction-centric animation, Animate Anyone 2 significantly advances the realism of AI-generated video. Despite this progress, challenges remain in scaling to very small objects and handling extreme proportion differences between the source and target characters. Future work will focus on refining intricate hand-object interactions and improving robustness for diverse character retargeting.
Summary updated