thu-ml/Causal-Forcing

Implements causal forcing—a diffusion distillation approach that conditions each generated frame on all previous frames rather than fixed intermediate frames—enabling frame-wise and chunk-wise architectures that unify text-to-video (T2V) and image-to-video (I2V) generation. Achieves real-time streaming video synthesis on consumer GPUs through autoregressive diffusion training followed by ODE initialization or consistency distillation, with native support for minute-level generation via Rolling Forcing integration. Built on the Wan diffusion backbone and compatible with CLIP embeddings, supporting both 1.3B and 14B parameter configurations.

456 stars.