Practical overview
When it fits
Use it when you want a self-hosted research codebase and documented checkpoints for experimenting with autoregressive video generation, including text- and image-conditioned inference.
Who it is for
- Researchers studying autoregressive diffusion distillation and causal consistency methods.
- Developers building self-hosted text- or image-conditioned video-generation prototypes.
Documented uses
- Explore Causal ODE and Causal Consistency Distillation as initialization methods for asymmetric DMD.
- Run documented text-to-video inference with chunk-wise or frame-wise checkpoints.
- Condition frame-wise video generation on an input image.
- Investigate minute-level long-video generation through the separate extension.
Requirements & limits
The repository is public and licensed under Apache-2.0. It is intended for self-hosted use; the documented environment uses Python 3.10 and Conda.
- The repository documents both chunk-wise and frame-wise models; frame-wise inference supports image conditioning.
- Causal Forcing++ includes documented one- and two-step frame-wise checkpoints, while the model list also includes four-step variants.
- The quick-start instructions require downloading Wan 2.1 model files and the relevant project checkpoint.
- The base model does not natively support videos longer than 81 frames; minute-level output is described as a separate Rolling Forcing-based extension.
Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.
Set up Causal Forcing for local inference
Prepare the documented Python environment and download the model and checkpoint files required by the inference workflow.
Before you start
- Conda
- Python 3.10
- Project repository
- Wan 2.1 model files and a Causal Forcing checkpoint
Steps
Create and activate the causal_forcing Conda environment with Python 3.10.
Install the repository requirements, OpenAI CLIP, and flash-attn, then install the project in development mode.
Download the Wan 2.1 model files and the Causal Forcing checkpoint needed for the chosen workflow.
Expected outcome
A locally prepared environment with model files and checkpoints for the repository’s documented inference stage.
Guide limitations
- This guide covers setup only; select a model configuration and follow the repository’s inference instructions to generate video.
- The base model does not natively support videos longer than 81 frames; minute-level generation uses a separate Rolling Forcing-based extension.
Sources & review
Source record details
- Source availability
- Active
- Sources checked
- 14 October 2026
- Original website title
- thu-ml/Causal-Forcing · Official homepage title · retained in its source language