FramePack-P1 is the next version of FramePack. The “P” means plan, prepare, prevision, plot, pre-arranging etc..

FramePack-P1 is based on FramePack with two new designs: Planned Anti-Drifting (d) and History Discretization (e).

Planned Anti-Drifting predicts sections that are far away from the next section before generating nearby sections. This reduces drifting “between” planned endpoints (frames will not drift between the endpoints).

History Discretization converts all history to discretization tokens (directly apply K-Mean to the entire dataset), aimed at finding a history representation that does not have obvious gap between train and inference. This reduces drifting “over” planned endpoints (the endpoints themselves will not drift). This is inspired by a potential observation that LLMs with discretization tokens tend to suffer less from drift compared to autoregressive video diffusion models.

The model uses 1 second as each section. The training data of this test are common clips/shots of about or less than 10 seconds. This is a challenging stress test that extrapolates beyond the training scope. We use relatively dynamic results to show that the model does not sacrifice motion dynamic range.

Videos compressed by h264crf18 for faster loading.

Videos compressed by h264crf18 for faster loading.

Each result iterates 12 promts, each promt for about 3.5 seconds: The man waves hands. The man laughs. The man talks. The man dances. The man waves hands. The man scratches his head. The man talks. The man dances. The man waves hands. The man spins around. The man talks. The man dances.

This web page is the first batch of results. More batches of results will be uploaded soon. The model and repo and paper will be updated soon.