Causal-Forcing

Overview of Causal-Forcing

What it is

Causal Forcing is an official codebase for the paper “Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation” and its extension Causal Forcing++. It introduces Causal ODE or Causal Consistency Distillation as a theoretically correct initialization for asymmetric DMD, enabling real-time interactive video generation. The project is developed by researchers from Tsinghua University, Shengshu, UT Austin, and RUC.

Key features

• Uses Causal ODE (Causal Forcing) or Causal Consistency Distillation (Causal Forcing++) to drive asymmetric DMD.
• Supports both chunk-wise and frame-wise models; the frame-wise model natively unifies text-to-video (T2V) and image-to-video (I2V) generation.
• Causal Forcing++ eliminates the need for ODE data curation and releases the first 1-step/2-step frame-wise models.
• Outperforms Self Forcing in visual quality and motion dynamics under the same training budget and inference efficiency.
• Provides checkpoints for chunk-wise 4-step, frame-wise 4-step, frame-wise 2-step, frame-wise 1-step, long video generation, and HY1.5-TI2V (8B) models.

Use cases

Real-time interactive video generation, long video generation (minute-level via Rolling Forcing), and as a base initialization for other video world models (e.g., DreamX-World 1.0, Matrix-Game 3.5, Astronex‑World 1.0, Zing-0.5, ZYT-World).

Getting started

bashconda create -n causal_forcing python=3.10 -yconda activate causal_forcingpip install -r requirements.txtpip install git+https://github.com/openai/CLIP.gitpip install flash-attn --no-build-isolationpython setup.py developDownload checkpoints and run CLI inference as described in the repository README.

License and pricing

Licensed under Apache-2.0. The project is open source with no stated pricing.

Based on the project's public README and repository metadata.