Causal-rCM: Open-Source Recipe for Streaming Video Generation
chenhsuanlin · x · 2026-08-21
Causal-rCM is an open-source algorithm and infra recipe for autoregressive video diffusion, designed for streaming video generation and interactive world models. It leverages the complementarity between forward (teacher-forcing) and reverse (self-forcing) divergences, combining Consistency Models (CM) and DMD.
Key Highlights:
- Wan2.1-1.3B achieves 84.63 VBench-T2V with 1-step / 2-step generation.
- Custom-mask FlashAttention-2 JVP kernel.
- First causal sCM / MeanFlow implementation for streaming video, offering 10× faster convergence.
- Full infra stack: FSDP2 / CP / SAC compatible with causal training and JVP.
- Reproducible inference, benchmark, and evaluation scripts.
The codebase serves as a tutorial for causal training paradigms and distillation algorithms.
More from Multimodal
- fal launches Worlds for real-time creation and interaction with AI-generated worlds — Scobleizer · 2026-10-16
- Gummy worms as a physics test: Gemini 4 Argon vs. Claude Opus 5.5 elasticity sims — DynamicWebPaige · 2026-10-16
- 260M-parameter looped diffusion transformer beats 6x larger baselines with 5x less compute — burkov · 2026-10-16
- Seedance 2.5 demo: stretching one second of an explosion into a 30-second shot — LudovicCreator · 2026-10-16
- Octopus Short Film Made with Grok Imagine, Upscaled via Topaz — creatoroff · 2026-10-16
- Open-sourced 61-second computing history film generated entirely by Muse Spark — sophiamyang · 2026-10-16