Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models

[remyx-ai]

Is your feature request related to a problem? Please describe.

Causal-rCM (arXiv:2606.25473) extends rCM-style diffusion distillation to autoregressive video diffusion. Its contributions are all training-time / weight-production: a teacher-forcing (TF, offline, forward-divergence, consistency-model-style) vs. self-forcing (SF, on-policy, reverse-divergence, DMD-style) paradigm split; the first teacher-forcing continuous-time consistency models (sCM / MeanFlow) for AR video, enabled by a custom-mask FlashAttention-2 JVP (Jacobian-vector-product) kernel (~10× faster convergence vs. discrete-time CMs); and a unified causal distillation recipe applied to Wan2.1-1.3B (2-step distilled, VBench-T2V 84.63) and Cosmos 3 (action-conditioned interactive world model), trained on synthetic data.

The problem: diffusers is inference-centric (pipelines, schedulers, models; training_utils.py is just lightweight helpers). Its only distillation-training presence is image-only LCM scripts (examples/consistency_distillation/) for SD/SDXL. There is no video distillation trainer, no DMD/MeanFlow/sCM, no causal teacher-forcing loop, and no custom-kernel path. The headline "2-step generation" is a property of the distilled weights, not a sampler/scheduler an inference pipeline consumes — so we need to decide whether/how to host an AR-video distillation trainer.

Describe the solution you'd like.

The realistic contribution is one of two scoped paths, depending on what the authors release:

• Inference-only slice (low-risk PR): if the authors' "open recipe" releases a Wan2.1-1.3B distilled checkpoint and it's convertible, ship a convert_* script + load the existing Wan2.1 pipeline at num_inference_steps=2 to demonstrate 2-step streaming generation. Confirm availability + license before committing.
• Distillation-training example: if we want to host an AR-video distillation trainer (precedent: examples/consistency_distillation/), scope a freestanding training script covering TF-consistency + SF-DMD + continuous-time CM, plus the custom-mask FlashAttention-2 JVP kernel.

Confirm the integration target is Wan2.1 / Cosmos 3 (the paper's actual models), not AnyFlow.

Describe alternatives you've considered.

• Skip the trainer, inference-only: consume only a released distilled checkpoint and avoid the trainer/kernel build entirely.
• Keep the FlashAttention-2 JVP / causal-mask kernel out-of-tree rather than taking a heavy native build in-repo.
• Note data/teacher constraints: teacher-forcing needs teacher weights and the paper trains on synthetic data — this may be out of scope for diffusers.
• The spec's original proposal to "integrate Causal-rCM into the AnyFlow pipeline" does not hold up: the paper operates on Wan2.1/Cosmos 3 (not AnyFlow), the method is a trainer, and AnyFlow/Wan/Cosmos3 ship as inference pipelines + transformers that would never call a distillation training method.

Additional context.

Why this is an Issue, not a PR. Causal-rCM's contribution is a distillation training recipe (TF consistency models + SF DMD + continuous-time sCM/MeanFlow, enabled by a custom FlashAttention-2 JVP kernel), not an inference-time algorithm. With no selection rationale naming an implementable subset and no released diffusers-format checkpoint, the realistic implementation would be a freestanding training module that the existing AnyFlow/Wan pipelines never call — so this needs maintainer discussion rather than a direct PR. The continuous-time-CM results hinge on a custom FlashAttention-2 JVP kernel plus AR-specific causal masking; none of this exists in the repo and it is a substantial standalone build, not a wiring task.

Open questions to unblock a PR:

1. Scope — host an AR-video distillation training example here, or inference-only (consume a released checkpoint) and skip the trainer?
2. Checkpoint — does the "open recipe" release a Wan2.1-1.3B distilled checkpoint? If yes and convertible, the low-risk PR is convert_* + Wan pipeline at num_inference_steps=2.
3. Kernel — is the custom FlashAttention-2 JVP / causal-mask kernel in scope, or out-of-tree?
4. Teacher + data — is there a teacher-weights / synthetic-data story we can support, or is this out of scope?

License & code availability: 🟡 no code repository or model URL surfaced and no LICENSE could be fetched (class no-code-link, compat 0.30) — worth confirming the paper has an open release before investing in adoption.

Drafted by Outrider — paper: arXiv:2606.25473.

Discovery context

Recommended paper: Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models
Confidence: high (Remyx relevance 1.00)
Research interest: diffusers

TL;DR (discovery loop): Causal-rCM is a video distillation TRAINING recipe (TF-consistency + SF-DMD + continuous-time CM via a custom FlashAttention JVP kernel), not an inference algorithm; diffusers is inference-only with just image-LCM training examples, so the orchestrator flagged a decision on whether/how to host an AR-video distillation trainer. Pre-flight routed this paper to Issue before the coding agent ran.

Reopen this Issue if you want Outrider to revisit this paper later. While it stays closed, the orchestrator will not re-recommend the same paper.