Causal Forcing++¶

Development line: project:causal-forcing · thread causal-forcing
Last event: 2026-05-16 · 1 dated since 2026-05-16 · Researched: 2026-09-05 · confidence: high

What it is¶

Causal Forcing++ is a research pipeline and frame-wise model release for low-latency text-to-video and image-to-video generation.

• Causal consistency distillation: replaces precomputed causal-ODE trajectories before asymmetric DMD.
• Reference checkpoints: provides 1-step and 2-step frame-wise models.
• Benchmark gains: outpaces 4-step chunk-wise Causal Forcing in paper benchmarks, but not as a production guarantee.

Development line¶

• 2026-05-16 — Causal Forcing++ public project resources were recorded. On 2026-05-16, we recorded project links to the site, source repository, and Hugging Face repository. The links mark public project resources, but confirm no specific version, technical result, or usage workflow.

What changed¶

2026-05-16 — Causal Forcing++ shipped causal consistency distillation with open 1-step and 2-step frame-wise checkpoints. Repository history dates the release to 2026-05-15, after paper v1 was submitted on 2026-05-14.

How to use this¶

From 2026-05-16, practitioners can evaluate Causal Forcing++ through linked project, source, and model locations. The dated record alone warrants no specific deployment or performance claims.

1. Create a Python 3.10 environment, then install requirements, CLIP, FlashAttention, and the package in editable mode. — https://github.com/thu-ml/Causal-Forcing
2. Download a Wan2.1 base model and either causal-forcing++/framewise-1step.pt or framewise-2step.pt from the model repository. — https://github.com/thu-ml/Causal-Forcing
3. Run inference.py with matching 1-step or 2-step frame-wise configs and --use_ema. Use the frame-wise path with an initial image for I2V. — https://github.com/thu-ml/Causal-Forcing

Best practices¶

• Test inference with the released checkpoints before running the three-stage training pipeline. — https://github.com/thu-ml/Causal-Forcing
• Switch training from BF16 to FP32 if causal-CD or DMD diverges or leaves frames blurry. — https://github.com/thu-ml/Causal-Forcing
• Do not use the base model past 81 frames; run Rolling Forcing for minute-scale video. — https://github.com/thu-ml/Causal-Forcing

Superseded by this¶

• 2026-05-15 — Causal Forcing++ replaces the causal-ODE Stage 2 step when initializing few-step models without saving paired ODE trajectories.
• 2026-05-15 — 1-step and 2-step frame-wise checkpoints replace the premise that Causal Forcing runs only as a 4-step chunk-wise model.

Still unknown¶

• Separate event logs and structured event feeds are not provided for the release.
• Official documentation covers only the research code and checkpoints, with no verified production deployment guides or hardware requirements.

Sources¶

Agent brief¶

• Subject: project:causal-forcing, thread causal-forcing, 1 dated events 2026-05-16 → 2026-05-16.
• Practical note: From 2026-05-16, practitioners can evaluate Causal Forcing++ through linked project, source, and model locations; the dated record alone justifies no specific implementation or performance claims.
• Confidence: high. Dated supersedes above mark what is obsolete.