Hugging Face Proposes ID-Forcing Framework: Revolutionizing Distribution Alignment in Long Video Generation
By Mr.Xu
Published:
Summary:Hugging Face's research team introduces the In-Distribution Forcing (ID-Forcing) framework to address the distribution drift problem in long video generation. While existing methods rely on key-value (KV) conditioning to mitigate drift, they overlook the issue of KV entries deviating from the training distribution during rollout. ID-Forcing ensures both KV caching and conditioning align with training configurations through a self-caching mechanism, significantly reducing drift in long video gene
Background and Challenges
Modern autoregressive (AR) video diffusion models excel at short-horizon video generation but face challenges in generating long videos due to drifting, where colors and textures shift, and motion dynamics decay. Existing methods primarily rely on key-value (KV) conditioning to mitigate drift by selecting or modifying cached KV entries. However, these methods assume that cached KV entries remain within the training distribution, an assumption that fails beyond the training horizon, leading to the KV-provenance problem where cached entries themselves become out-of-distribution (OOD).
ID-Forcing Solution
To address this, Hugging Face's research team proposes the In-Distribution Forcing (ID-Forcing) framework. This framework ensures that both KV caching and conditioning align with training configurations through the following mechanisms:
- Self-Caching Mechanism: Each chunk is cached without attending to prior KV entries, keeping the rolling window exactly in-distribution.
- KV Conditioning Alignment: Adjusts KV conditioning records at test time to match training configurations.
Technical Highlights
- Innovative Self-Caching Mechanism: Prevents OOD KV entries at their source.
- Seamless Extension: Enables short-horizon models to generate minute-scale videos.
- Superior Performance: Demonstrates competitive performance on standard benchmarks and substantial improvements in mitigating drift as validated by user studies.
Experiments and Validation
The research team conducted extensive evaluations, demonstrating that ID-Forcing significantly outperforms existing methods in mitigating drift while maintaining competitiveness in standard video generation tasks. User studies further validate the method's effectiveness, showcasing its potential in practical applications.
Industry Impact and Future Directions
The introduction of the ID-Forcing framework provides a new technical path for the long video generation field, promising to drive further advancements in video generation technology. Its innovative approach to handling distribution drift also offers insights for AI models in other domains. In the future, the research team plans to apply this technology to a wider range of video generation tasks and explore its potential in other areas.
— END —Source: Hugging Face Daily Papers (2026-10-02)
Tags: #Hugging Face #Video Generation #Long Video #AI Models #ID-Forcing
Community Comments