thu-ml/Causal-Forcing

What it solves

Causal Forcing and Causal Forcing++ address the challenge of generating high-quality, real-time interactive videos. They improve upon previous autoregressive (AR) diffusion models by enhancing visual quality and motion dynamics while maintaining inference efficiency and training budgets.

How it works

The project employs a multi-stage training pipeline to create a theoretically correct initialization for asymmetric Distribution Matching Distillation (DMD):

1. Stage 1: AR Diffusion Training: Establishes the base autoregressive diffusion model.
2. Stage 2: Initialization: Uses either Causal ODE (Causal Forcing) or Causal Consistency Distillation (Causal Forcing++). The latter eliminates the need for time-consuming ODE data curation by using only ground-truth data.
3. Stage 3: Asymmetric DMD: The final stage that produces the few-step models used for generation.

It supports both chunk-wise and frame-wise models, with the frame-wise approach natively unifying text-to-video (T2V) and image-to-video (I2V) generation.

Who it’s for

Researchers and developers working on video generation, world models, and real-time interactive video systems who need high-quality motion and low-latency inference.

Highlights

• Extreme Efficiency: Offers 1-step and 2-step frame-wise models for extremely low latency.
• Unified Generation: Frame-wise models natively support both T2V and I2V.
• Scalability: Compatible with extensions like Rolling Forcing for minute-level long video generation.
• Industry Adoption: Used as a base for several SOTA video world models including DreamX-World 1.0 and Zing-0.5.