Open-source project
thu-ml/Causal-Forcing avatar
thu-ml/Causal-Forcing

Causal Forcing: five checkpoint files, 36 dependency lines, and four exact pins

[ICML 2026] Official codebase for "Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation" & Causal Forcing++

1,026 stars62 forksPythonApache-2.0

At a glance

What is it?
thu-ml/Causal-Forcing is the code release behind the Causal Forcing and Causal Forcing++ papers on real-time interactive video generation, built on asymmetric DMD initialization and shipped as checkpoint files on a model hub rather than as GitHub releases. The packaging stub declares no dependencies at all, and the quick start stops just short of the installation section it advertises.
Who is it for?
This is a research code drop with the shape of one: the method description, the training stages and the checkpoint list are all here, and everything a newcomer needs in order to reproduce a run is either inherited from another project or pointed at off-repository. Judge it on that basis.
Can I use it commercially?
Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository last received commits 22 days ago.
What is it written in?
Mainly Python, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 13, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • thu-ml/Causal-Forcing1,026 ★ · +69
Star history
Repository2026-09-112026-11-13
thu-ml/Causal-Forcing9571k

69 Stars in the last 63 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

setup.py declares version 0.0.2 and no dependencies

The only packaging file in the tree is setup.py, and it is three fields long:

python
from setuptools import setup, find_packages
setup(
    name="causal_forcing",
    version="0.0.2",
    packages=find_packages(),
)

There is no install_requires, no extras_require and no console entry point. Every dependency lives in a separate requirements.txt at the root, which setup.py never reads, so an install performed through the packaging metadata arrives with none of the libraries the inference path needs. The distribution name uses an underscore where the repository name uses a hyphen, and find_packages() decides which directories ship rather than an explicit list, across a root that holds model/, pipeline/, utils/, trainer/, prompts/, demo_utils/, wan/ and long_video/. The version number says something about release discipline too: two papers, an ICML 2026 badge in the description, and a package that reports itself as 0.0.2.

Four exact pins inside 36 dependency lines

requirements.txt is mostly open ended, with four exceptions written in the double equals form:

bash
diffusers==0.31.0
numpy==1.24.4
av==13.1.0
pydantic==2.10.6

Everything next to them is a floor rather than a version:

bash
torch>=2.4.0
torchvision>=0.19.0
transformers>=4.49.0
accelerate>=1.1.1

So the four nails hold the diffusion library, the array library, the video decoder and the validation library, while the torch family stays open and twenty six of the thirty six lines carry no version at all. One group in the list has nothing to do with training: onnx, onnxruntime, onnxscript, onnxconverter_common and dominate, then flask, flask-socketio and starlette, then pycocotools, lmdb, matplotlib and scikit-image. Export, serving and measurement all sit in one file with the core stack, and the quick start stops before the installation section, so the page never narrows the list down to what one inference run actually needs.

The inference environment is claimed identical to another project

The quick start opens with a single sentence: the inference environment is identical to Self Forcing. Nothing follows that describes an environment. What follows is a ceiling on output length. As with CausVid and Self Forcing, Causal Forcing does not natively support videos longer than 81 frames, and as a base training method it is orthogonal to techniques such as Longlive and Rolling Forcing, with a pointer to a long video extension further down for anyone who wants it as a baseline.

That ordering sets expectations oddly. The claim of identity is doing the work of an installation guide, the one hard number attached to the method itself is 81 frames, and that number is a limit rather than a capability. The quick start ends at that pointer. The installation entry and the three numbered training stages in the table of contents sit below it, where the page stops.

Three ODE data collection scripts sit at the root after the second paper removed that step

Causal Forcing++ replaces the ODE stage with causal Consistency Distillation in order to eliminate ODE data curation, and the root of the tree still carries three curation scripts side by side: get_causal_ode_data_chunkwise.py, get_causal_ode_data_framewise.py and get_causal_ode_data_kv_optimized.py, the last of which is named for the cache it optimises rather than the stage it serves. Those scripts belong to the first method, and the news log shows the stage being worked on rather than retired. An April entry reports the stage 2 consistency distillation infrastructure optimised for three times faster training, with a checkpoint released alongside it. A May entry credits a contributor with making the chunk-wise ODE data curation three times faster.

So the removal is a property of the second method, not of the repository. Anyone running stage 2 the ODE way is still assembling paired data with three scripts at the root, and anyone running it the Causal Forcing++ way is not.

Five checkpoint files, seven table rows, and two rows belonging to another repository

The checkpoint table has seven rows, and five of them point at files inside this project's own collection: chunkwise/causal_forcing.pt for the chunk-wise 4 step model, described as an autoregressive model on Wan1.3B outperforming Self Forcing; framewise/causal_forcing.pt for the frame-wise 4 step model; framewise-2step.pt and framewise-1step.pt; and chunkwise/longvideo.pt for a minute-long generator. The other two rows, for the HY1.5-TI2V 8B model and for the action-conditioned world model, point at a different collection and tell you to refer to shengshu-ai/minWM for the details.

The folder names carry the version split. The two four step files sit directly under chunkwise and framewise with no suffix, while the one and two step files sit under a directory whose name is the project name with the two plus signs percent encoded. If you are matching a file to a paper, the directory is the only marker. The two step row is even annotated as better than the four step row sitting two lines above it, with no metric attached.

No GitHub releases, so the model files are the only published artifacts

The repository has no GitHub releases, and the package version is 0.0.2, so nothing here is versioned on a releases page. The artifacts a user fetches are model files on a model hub, and the change history is a dated news list on the page instead. That list runs from the 2 February release of paper, project page and code, through consistency distillation as a preview in early February, Rolling Forcing support for minute-level video, image to video support, an added FAQ section, adoption by three other projects in March, the faster training work in April, the second paper and an 8B model release in May, two more projects building on the initialization in July, and a contributed ComfyUI node in October.

The comparison to Self Forcing is made in one sentence, in both visual quality and motion dynamics at the same training budget and inference efficiency. No metric, no evaluation set and no table accompany it on the page, so the sentence is the claim and the number behind it lives in the paper.

Seven projects claim this initialization, and none of them runs these checkpoints

Adoption is the strongest signal on the page, and it is catalogued in two directions. Five recent state of the art video world models are named as built on Causal Forcing or Causal Forcing++ initialization: DreamX-World 1.0, Matrix-Game 3.5, Astronex-World 1.0, Zing-0.5 and ZYT-World. A July entry adds RAVEN and Self Gradient Forcing as built on the Causal Forcing initialization. A March entry lists Rolling Sink, Infinity-RoPE and Deep Forcing as adopting Causal Forcing as one of their base models.

Read the direction of those claims carefully. Every one of them is about initialization, meaning other projects begin their own training from this one, and none of them is about running these checkpoints or reusing this inference code. The newest entry, dated 1 October, credits a contributor with a ComfyUI node for CF++ that lives in someone else's repository, which is the only mention of ComfyUI on the page and is a pointer outward rather than code here.

No paper in the tree, and a Feishu document as the only written guidance

Nothing in the tree carries the paper. There is no PDF, no citation file and no bibliography entry; the papers are reached through two arXiv links, the project through a GitHub Pages site, and the only prose document linked is a Feishu wiki page that arrives with a copy tracking query attached to the address. The code itself is Apache-2.0, with a LICENSE at the root next to assets and a prompts directory and the training and inference scripts.

The counts sit oddly beside that. The repository has 985 stars, 56 forks and 33 open issues, with the last commit on 2026-09-24 and no releases to sort them into. The page also carries two small slips in the places a newcomer reads first: the table of contents renders one training entry as Causal Foricng, and the May entry for the second paper spells the method Casual Consistency Distillation. Neither changes what runs.

Editorial conclusion

This is a research code drop with the shape of one: the method description, the training stages and the checkpoint list are all here, and everything a newcomer needs in order to reproduce a run is either inherited from another project or pointed at off-repository. Judge it on that basis. What you get in the tree is inference and training scripts, a stage 2 that comes in an ODE flavour and a consistency distillation flavour, and curation scripts for the first flavour only. What you have to supply is the environment, since setup.py declares no dependencies and the one hard statement about it is that it matches another project, and the honest reading of the quality claim is that it lives in the paper rather than on the page. Before trusting a checkpoint, check which paper its directory came from, since the second paper's files sit in a differently named folder, and before trusting a rebuild, note that twenty six of the thirty six dependency lines carry no version at all.

Frequently asked questions

What is thu-ml/Causal-Forcing?

The official code release for Causal Forcing and Causal Forcing++, the pair of papers on real-time interactive video generation from Tsinghua University with co-authors at Shengshu, UT Austin and Renmin University. Both use asymmetric DMD, initialized from a causal ODE in the first paper and from a causal Consistency Distillation stage in the second.

How long are the videos Causal-Forcing can generate?

The page states a ceiling rather than a target: as with CausVid and Self Forcing, Causal Forcing does not natively support videos longer than 81 frames. Longer output is treated as an extension problem, with Rolling Forcing supported for minute-level generation and a separate long video extension pointed at for use as a baseline.

Where are the model weights for Causal-Forcing?

On a model hub collection, not in GitHub releases, since this repository publishes none. Five files are listed: chunkwise and framewise causal_forcing.pt, framewise-2step.pt and framewise-1step.pt from the second paper's folder, and chunkwise/longvideo.pt for a minute-long generator.

Does Causal-Forcing release code only, or checkpoints too?

The repository itself holds the code. Checkpoints live on a model hub, and two of the seven rows in the checkpoint table point at a different project's collection for the HY1.5-TI2V 8B model and the action-conditioned world model, both of which redirect you to shengshu-ai/minWM.

How do I install Causal-Forcing?

The quick start states that the inference environment is identical to Self Forcing rather than spelling one out. In the tree, setup.py declares the name causal_forcing, version 0.0.2 and find_packages() with no install_requires, and the 36 line requirements.txt at the root is a separate file that setup.py does not read.

Official sources

  1. Issues
  2. License: Apache-2.0
  3. Project website
  4. README
  5. thu-ml/Causal-Forcing on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/thu-ml-causal-forcing.svg)](https://hysenlabs.com/projects/thu-ml-causal-forcing)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/thu-ml-causal-forcing.svg)](https://hysenlabs.com/projects/thu-ml-causal-forcing)

Was this useful?