Generating long AI videos is hard. Models are trained on short clips and build videos chunk by chunk. Each chunk goes through many refinement steps, but existing methods discard all intermediate predictions. This throws away useful context and lets errors accumulate.
Trajectory Forcing preserves the full history of these predictions as context for the next chunk. It learns this by first simulating the steps on training data, then using its own predictions during a practice phase.
This technique achieved the highest quality score of 85.15 on a standard benchmark. For videos six times longer than its training data, it showed the best quality and lowest drift, proving that using the full generation history stabilizes long-form video.