Wan2.1

Headlines · 3

1. 3d
   AB-Attention accelerates video diffusion transformers

   Bin Xu and eight other authors submitted “AB-Attention: Accurate Block Attention for Accelerating Video Diffusion Transformers” to arXiv on October 29, 2026. The training-free method combines block selection, inclusion-probability weighting and finer key/value gathering to approximate dense attention at high sparsity. On Wan2.1 and HunyuanVideo, the authors report skipping 89–92% of attention computation and achieving up to 2.23× end-to-end speedup.

   Hugging Face · Papers · 🔥 0
2. Oct 20
   ALLOT: training-free block-sparse attention allocation for video DiTs

   A new arXiv preprint by Tianze Xia and five co-authors proposes ALLOT (Allocating Layer-step Latency by Output Tests), a training-free method for allocating block-sparse attention in video Diffusion Transformers. The authors observe that sparsifying attention in different layers and denoising steps affects final-video fidelity to very different degrees, while the most sensitive layer-step pairs stay nearly the same across inputs. ALLOT probes each layer-step cell and calibrates along the denoising trajectory, then jointly optimizes cell densities within a measured latency budget; the paper reports an 11.8 dB PSNR gain over XAttention at equal or shorter wall-clock time on HunyuanVideo, and wins or ties the default in all nine settings across Wan2.1, Wan2.2 and HunyuanVideo.

   Hugging Face · Papers · 🔥 0
3. Oct 6
   GRACE: generation-aware latent compression cuts Wan2.1 video diffusion tokens 8x

   An arXiv paper introduces GRACE (Generation-Aware Latent Compression for Efficient Video Generation), a two-stage framework that compresses a pretrained video autoencoder while keeping it compatible with a pretrained DiT: it keeps a frozen base latent, learns a residual latent for information lost under stronger compression, and aligns the compressed latent in the frozen DiT's feature space so the autoencoder is optimized for generation. The DiT is then adapted with lightweight fine-tuning and asymmetric denoising in which the base is denoised ahead of the residual. The authors report an 8x token reduction and 11.1x latency reduction for Wan2.1-I2V-14B at 480x832x81 while matching the pre-compression pipeline's quality on VBench.

   Hugging Face · Papers · 🔥 0

Experience and discussion from the community

Share my Wan2.1 experienceAsk about Wan2.1