Introduction

Wan already makes gorgeous video—until the waiting starts. Today Alibaba Research and collaborators open SparkDiffusion, a post-training stack that aims to keep that look while cutting the cost of long spatiotemporal attention. Same-day Hugging Face drops under the SparkWan name put sparse, few-step Wan checkpoints in creators’ hands.

What’s New

SparkDiffusion is framed as a unified acceleration recipe for visual generation: a short sparse warm-up, few-step trajectory-mixed distillation (CrossDistill), and FP8 fused kernels—built to dodge what the paper calls the high-sparsity trap, where step-local loss looks fine but terminal video quality stalls once attention gets extremely sparse.

The default sparse module is RoLA (rotary-positioned low-rank linear attention): a high-energy sparse branch plus a lightweight low-rank compensation path so global context doesn’t vanish when you keep only a tiny fraction of query–key blocks.

On the paper’s headline setting—Wan2.1-T2V-14B-720P, 3-step CFG-free inference at high sparsity with FP8—the authors report up to about 265× end-to-end speedup on a single RTX 5090 versus a 50-step CFG dense baseline (about 220× on H100). Compact Wan2.1-T2V-1.3B-480P is clocked near 1.3 s end-to-end at 90% sparsity on that same consumer card.

Coverage spans Wan2.1 and MoE-style Wan2.2, T2V and I2V, and 480P/720P—so this is less “one hacky LoRA” and more a family of accelerated backbones.

SparkWan Weights Hit Hugging Face

Alongside the arXiv drop, alibabagroup published SparkWan weights on Hugging Face on 2026-09-24 (UTC morning). The Wan2.2 spotlight card is:

alibabagroup/SparkWan2.2-T2V-14B-480P-0.95Sparsity

It ships high-noise and low-noise expert checkpoints for Wan2.2 T2V A14B at 480P, Apache-2.0, base Wan-AI/Wan2.2-T2V-A14B, 4 inference steps, and supported attention sparsity roughly 90–95% (topk 0.1–0.05).

Same-day siblings cover SparkWan2.1 T2V/I2V sparsity variants (including 720P 0.97 sparsity and a 3-step 0.95 pack)—handy if you’re already on Wan2.1 runtimes.

Code and docs live at the project site and GitHub (AlibabaResearch/SparkDiffusion, ~113★, pushed 2026-09-23), with the main paper at arXiv 2609.23153 (plus companion RoLA 2609.06712 and CrossDistill 2609.14725).

Why Creators Should Care

If you’ve been eyeing Wan locally but flinching at DiT attention cost, sparse + few-step + FP8 is the combo that actually moves the needle—when quality holds. SparkDiffusion’s pitch is that extreme sparsity is usable if you stage the training: adapt the sparse architecture first, then terminal-align with trajectory-mixed distillation instead of hoping step-local flow matching alone will save the final frames.

Practical takeaway: start from the released SparkWan experts that match your Wan major version, keep sparsity in the supported band, and treat the few-step schedule as first-class—not an afterthought. Benchmark on your own prompts; the paper’s speedups are system results (steps × sparsity × FP8), not magic from any single knob.

(Same-day sidenote for image folks: Alibaba PAI also posted official 4-NFE PDD Acc LoRAs for Qwen-Image-2.1—a different lane, but the “few-NFE acceleration” mood is clearly in the air.)

Links

Original Source

https://github.com/AlibabaResearch/SparkDiffusion

Conclusion

Open sparse Wan acceleration just got a lot more concrete: a readable recipe, companion papers, and downloadable SparkWan experts on day one. If you’re shipping Wan locally, this is the drop to put on the weekend bench.

—Aurelia ♡