An acceleration framework for Diffusion Transformer (DiT) video generation, released from Alibaba's research org and built around a diagnosis the authors call the high-sparsity trap: pushed past roughly 90% attention sparsity, step-local training loss keeps falling while the finished videos still show broken structure, semantic drift, mosaic textures and temporal flicker. The paper argues the failure is one of supervision rather than capacity — the dominant terminal errors originate in the high-noise structure-generation stage, which extended step-local training cannot fix — and turns that into a staging principle: first adapt the sparse architecture into a coarse prior with a short sparse warm-up, then correct the terminal distribution with few-step trajectory-mixed distillation. The shipped stack composes three multiplicative pieces: RoLA rotary-positioned low-rank linear attention (arXiv 2609.06712), CrossDistill trajectory-level hybrid few-step distillation (arXiv 2609.14725), and W8A8 FP8 quantization with self-developed fused kernels that need no external sparse-attention checkout at inference time.

The headline 265× is a specific, bounded measurement, not a general speedup. It is the end-to-end diffusion-latency ratio for Wan2.1-T2V-14B at 720×1280, 81 frames, on a single RTX 5090, comparing 3-step CFG-free inference at 97% attention sparsity with FP8 fused kernels (NFE 3) against a 50-step classifier-free-guidance BF16 dense baseline running FlashAttention-2 (NFE 100): 4,769 s down to 18.0 s. The same comparison on an H100, whose baseline uses FlashAttention-3, gives 220× — 1,757 s to 8.0 s. Text encoding and VAE decoding are excluded from both numbers, and the isolated 90%→97% sparsity stretch accounts for only 1.31–1.32× of the total. At 97% sparsity quality reads VBench 83.15 / VBench-2.0 58.05 against the dense model's 83.69 / 60.20 and the strongest 90%-sparsity baselines TurboDiffusion (82.88 / 57.98 at 25.3 s and 16.0 s) and FastWan with VSA (82.72 / 58.04 at 54.1 s and 20.5 s). On the compact Wan2.1-T2V-1.3B-480P at 90% sparsity it takes the best VBench-2.0 of the accelerated methods (55.85 against 54.63 and 54.61) at the lowest latency, 1.3 s on RTX 5090 and 0.6 s on H100. A seed-diversity check over 1,000 prompts with V-JEPA 2 and VideoMAE V2 encoders puts it closest to the dense reference among the few-step methods.

Apache 2.0, Python, about 400 GitHub stars. Six checkpoints covering Wan2.1 T2V/I2V at 480P and 720P and Wan2.2 T2V at 480P are in the HuggingFace collection, at 0.90–0.97 sparsity and 3–4 steps; the repository provides sparse-finetuning, distillation and single-case inference launchers, and credits NVIDIA's rCM for the distillation implementation and Tsinghua's SLA kernel, still required for RoLa sparse training via SLA_SRC. Of 15 authors, 8 list Alibaba Group (including corresponding author Tengxu Sun) and 6 list Peking University (including corresponding authors Kun Yuan and Bin Cui), with Tsinghua University, UESTC and Harbin Institute of Technology also represented. The date marks the repository going public; the paper was posted 2026-09-19 and the weights followed on 2026-09-25.

Paper

Authors: Yuxi Liu · Haoyu Li · Zekun Zhang · Tengxu Sun · Yixiang Cai · Jiayong Li · Yifei Xia · Tianle Liu

Library

Language Python
License Apache 2.0
videodiffusionefficiencyinferenceattentionsparsequantizationdistillationopen-source