Causal Forcing
[ICML 2026] Official codebase for "Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation" & Causal Forcing++
Last verified:
What is Causal Forcing?
Causal Forcing is a novel method for high-quality real-time interactive video generation that addresses a theoretical flaw in Self Forcing's training pipeline. It uses an autoregressive (AR) teacher for ODE initialization, thereby bridging the architectural gap that occurs when distilling pretrained bidirectional video diffusion models into few-step AR models. The method first fine-tunes a bidirectional base model into an autoregressive diffusion model, then uses it as the teacher for ODE initialization, followed by the same DMD stage as in Self Forcing.
Key features include significantly outperforming Self Forcing in both visual quality and motion dynamics while maintaining the same training budget and inference efficiency. The method surpasses SOTA Self Forcing by 19.3% in Dynamic Degree, 8.7% in VisionReward, and 16.7% in Instruction Following. It enables real-time streaming video generation on a single RTX 4090 and releases both frame-wise and chunk-wise models, with the frame-wise setting being a configuration Self Forcing chose not to release.
Causal Forcing is designed for researchers and developers working on video generation, particularly those interested in real-time interactive video applications. It is built on top of open-source implementations of CausVid, Self Forcing, and the Wan2.1 repo, and is compatible with Self Forcing training configs and checkpoints for seamless migration.
Causal Forcing pricing
Pricing model: Freemium
Free and open source under Apache-2.0 license. The codebase, pretrained checkpoints, and models are freely available on Hugging Face and GitHub. No paid tiers or commercial licensing required.
Causal Forcing pros
- Significantly outperforms Self Forcing in visual quality
- Significantly outperforms Self Forcing in motion dynamics
- Same training budget as Self Forcing
- Same inference efficiency as Self Forcing
- Enables real-time streaming video generation
- Runs on a single RTX 4090
- 19.3% improvement in Dynamic Degree over Self Forcing
- 8.7% improvement in VisionReward over Self Forcing
- 16.7% improvement in Instruction Following over Self Forcing
- Offers both frame-wise and chunk-wise models
- Frame-wise model provides higher dynamic degree and more expressive output
- Chunk-wise model offers more stable generation
- Seamless migration from Self Forcing configs and checkpoints
- Full-stack open source code available
- Uses AR teacher for theoretically correct ODE initialization
Causal Forcing cons
- Requires multi-node distributed training with 8 nodes and 8 GPUs per node
- Requires flash-attn installation which can be complex
- Needs pretrained Wan2.1 models (1.3B and 14B variants)
- Frame-wise config not compatible with Self Forcing framework without migration
- Requires Python 3.10 specifically
- Training requires downloading a 6K toy dataset first
- Inference environment must match Self Forcing exactly
- Only supports text-to-video generation based on Wan2.1 base models
Frequently asked questions about Causal Forcing
What is Causal Forcing?
Causal Forcing is a novel method for autoregressive diffusion distillation that addresses the architectural gap in distilling bidirectional video diffusion models into autoregressive models. It uses an AR teacher for ODE initialization, thereby bridging the theoretical gap and significantly improving video generation performance compared to Self Forcing.
How does Causal Forcing differ from Self Forcing?
Causal Forcing identifies a theoretical flaw in Self Forcing's ODE initialization: using a bidirectional teacher to supervise an autoregressive student violates frame-level injectivity. Causal Forcing first fine-tunes a bidirectional base model into an autoregressive diffusion model, then uses it as the teacher for ODE initialization, which is theoretically correct.
What performance improvements does Causal Forcing offer?
Causal Forcing surpasses SOTA Self Forcing by 19.3% in Dynamic Degree, 8.7% in VisionReward, and 16.7% in Instruction Following, while outperforming all baselines across all metrics.
What hardware is required to run Causal Forcing?
Causal Forcing enables real-time streaming video generation on a single RTX 4090 for inference. For training, it requires multi-node distributed setup with torchrun using 8 nodes and 8 GPUs per node.
What base models does Causal Forcing use?
Causal Forcing uses Wan-AI/Wan2.1-T2V-1.3B and Wan-AI/Wan2.1-T2V-14B as base models, which must be downloaded from Hugging Face before use.
What are the frame-wise and chunk-wise models?
The frame-wise model provides higher dynamic degree and more expressive output, while the chunk-wise model is more stable. Self Forcing chose not to release the frame-wise setting, but Causal Forcing open-sources both.
Is Causal Forcing compatible with Self Forcing?
Yes, the inference environment is identical to Self Forcing, and Stage 3 (DMD) is compatible with Self Forcing training, allowing seamless migration using their configs and checkpoints.
What is the license for Causal Forcing?
Causal Forcing is released under the Apache-2.0 license, making it fully open source for both research and commercial use.
How do I install Causal Forcing?
Install by creating a conda environment with Python 3.10, then running: pip install -r requirements.txt, pip install git+https://github.com/openai/CLIP.git, pip install flash-attn --no-build-isolation, and python setup.py develop.
Can I skip the training stages?
Yes, pretrained checkpoints are available for both Stage 1 (AR-diffusion) and Stage 2 (Causal ODE Initialization), so you can download them directly from Hugging Face and skip training if desired.