Event
Fréchet Forcing Trains Long-Horizon World Models with Short Video Clips
1 reports · 1 sources
The authors of the Fréchet Forcing paper propose a fine-tuning method that applies a flow-matching loss where self-generated history meets real data, and uses Fréchet distance to constrain context fidelity and visual realism; all three constraints come from short video clips, requiring no teacher model, online judge, or long videos. The paper reports that after fine-tuning MiniWorld-1B, which was pretrained on 253-frame clips, on 381-frame clips for up to 36 GPU hours, the generations can stably exceed 4000 frames in two task types—robotic manipulation and camera-control exploration—and continue to follow action instructions.
Report timeline
- Fréchet Forcing: Training Long-Horizon Streaming World Models with Short Video Clips
Fréchet Forcing proposes a fine-tuning method that keeps streaming video world models stable during long-horizon generation beyond the training length and lets them continue to follow action commands. The method applies a flow-matching loss where self-generated history meets real data, constrains context fidelity with a Fréchet distance centered on the real clip itself, and enforces visual realism with a Fréchet distance on real frames in a frozen perceptual space. All three constraints come from short video clips, with no teacher model, online judge, or long videos needed.
HuggingFace Daily Papers ·