CausalWan2.2-I2V-A14B-Preview — FastVideo causal I2V model on Wan2.2-A14B

TL;DR

CausalWan2.2-I2V-A14B-Preview is a Hugging Face model release from Hao AI Lab’s FastVideo team — a causal/autoregressive image-to-video distillation of the Wan2.2-A14B MoE I2V base, packaged in the Diffusers format. The card markets it as part of the FastWan collection (“models trained with video sparse attention and distillation”), porting the Self-Forcing causal-AR + sparse-distillation recipe from the original Wan2.1-T2V-1.3B / Wan2.1-T2V-14B targets to the much larger Wan2.2-A14B mixture-of-experts I2V backbone. The card explicitly flags it as a preview with “quality issues” and “unoptimized inference speed” — the artifact value is the public availability of a causal Wan2.2-A14B I2V checkpoint that the wiki’s existing CausVid / Self-Forcing / Causal Forcing / FastWan lineage can be evaluated against, not a finished SOTA claim.

Key claims

The model is a Diffusers-format I2V checkpoint derived from Wan2.2-A14B and trained inside the FastVideo “models trained with video sparse attention (VSA) and distillation” collection [model card header, “FastWan Collection” link].
The card explicitly markets this as a preview with known quality regressions and unoptimized inference — i.e. not yet a published SOTA operating point [card body, “Note that this is a preview model, meaning there are still quality issues. The inference speed is also unoptimized.”].
VSA (the sparse-attention kernel underlying the FastWan recipe) is identified by the card as arxiv.org/abs/2505.13389 — same kernel the team’s FastWan blog co-trains with DMD on Wan2.1 [collection description].
A sibling SFWan2.2-I2V-A14B-Preview-Diffusers checkpoint exists and is the one demonstrated in the FastVideo inference recipe — the inference snippet uses an 8-step DMD denoising schedule [1000, 850, 700, 550, 350, 275, 200, 125] at 81 frames / 832×480 [card body, FastVideo Python example].
Inference is structured for the Wan2.2 MoE: dit_cpu_offload=True is set in the FastVideo example with the comment “DiT need to be offloaded for MoE”, and DiT precision is forced to fp32 while text encoder + VAE offload separately — i.e. the recipe assumes consumer-class memory, not a full MoE resident on a single GPU [card body, FastVideo Python example].
The Diffusers usage path is the standard DiffusionPipeline.from_pretrained(...).to("cuda") flow at dtype=torch.bfloat16 and runs end-to-end via pipe(image=image, prompt=prompt).frames[0] → export_to_video [card body, Diffusers usage].

Method

The HF card itself does not describe a method; the artifact reuses the FastVideo team’s published recipe and points to it. From the collection metadata and the related SFWan2.2-I2V-A14B-Preview-Diffusers sibling, this preview is the causal-attention Wan2.2-A14B I2V variant (Self-Forcing-style chunkwise AR rollout with KV caching, the “Causal” prefix in the model name), co-trained with Video Sparse Attention. The FastWan blog (FastWan: Generating a 5-Second Video in 5 Seconds via Sparse Distillation) is the canonical description of the recipe: VSA is added to the student only, DMD is the distillation objective, and the real-/fake-score networks remain full-attention.

The Wan2.2-A14B base is a mixture-of-experts I2V DiT — distinct from the Wan2.1-T2V-1.3B and 14B-dense targets previously demonstrated for FastWan. Two consequences are visible from the model card alone: (a) the inference example forces DiT CPU offload to fit the MoE on consumer hardware, and (b) fp32 is used for the DiT despite a bfloat16 pipeline dtype, suggesting routing or expert-weight sensitivity that the team has not yet resolved at low precision. The “preview / quality issues” disclaimer is consistent with this being the first FastWan-style port to an MoE I2V base.

There is no per-step benchmark, no VBench number, and no head-to-head comparison published on the card itself. The 8-step DMD schedule visible in the inference example is the only training-derived hyperparameter that leaks into the release; it matches the operating point family used by the FastVideo team in prior FastWan releases.

Results

The card publishes no quantitative results — no VBench, no FPS, no latency, no GPU-hours-to-train, no quality ablation. The only quantitative information visible:

Inference sequence: 81 frames at 832 × 480, 8 DMD denoising steps in the SFWan sibling example.
Recommended precision/offload recipe: bfloat16 pipeline, fp32 DiT, CPU-offload DiT + text encoder, VAE on GPU.

This release is a checkpoint announcement, not a paper. Concrete numbers will need to come from a future FastWan-style blog post or the team’s GitHub.

Why it’s interesting

The wiki’s Autoregressive Video Generation cluster has converged on a clear question (BUD.md §6: ≥3 papers = pattern): how to do causal AR distillation correctly on Wan-class video DiTs. The filed papers — Self Forcing (Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion), Causal Forcing (Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation), Causal Forcing++ (Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation) — all use Wan2.1-T2V-1.3B as the base because it’s the cheapest base where the recipe is verifiable. The interesting feature of this release is that FastVideo is now publicly shipping the same family of recipe on Wan2.2-A14B I2V — a mixture-of-experts image-to-video base, structurally different from the T2V-dense lineage everything else has been demonstrated on. That gives the cluster its first MoE / I2V data point, and the “preview, known quality issues” framing is itself useful evidence that the recipe doesn’t drop in trivially across base-model architectures.

It also closes the loop with FastWan: Generating a 5-Second Video in 5 Seconds via Sparse Distillation: FastWan explicitly noted that “Wan2.2-TI2V-5B’s ~20K sequence length is too short to benefit from sparse attention; only DMD with full attention is applied,” and stopped short of the larger Wan2.2-A14B MoE variant. This checkpoint is the next domino — sparse + causal-AR + DMD applied to the largest current Wan2.2 base — and is being released before the supporting blog/paper, so the wiki should treat it as a pointer for now and wait on numbers.