Research map / Video / Video generation
SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer
Junsong Chen, Yuyang Zhao, Jincheng YU, Ruihang Chu, Junyu Chen, Shuai Yang, Xianbang Wang, Yicheng Pan, Daquan Zhou, Huan Ling, Haozhe Liu, Hongwei Yi and 8 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace full attention with linear attention (ablation) | SANA-Video · 720p video generation | latency: 4x speedup at 720P | Improved |
| replace full attention with linear attention (ablation) | SANA-Video · 480p video generation | latency: 2x speedup at 480P | Improved |
| add 3D RoPE after ReLU in linear attention (ablation) | SANA-Video · training loss | training loss | Improved |
| add 1D temporal convolution to Mix-FFN (ablation) | SANA-Video · training loss and motion | training loss | Improved |
| use monotonically increasing SNR timestep sampler instead of random (ablation) | LongSANA · autoregressive block training | video quality and block consistency | Improved |
| remove RoPE from key or query in the denominator of linear attention (ablation) | SANA-Video · training stability | training stability | Improved |
| use DCAE-V instead of Wan2.1VAE under latent noise 0.2 | VAE reconstruction · Panda-70M 1000 samples | PSNR: 24.25 -> 29.34 | Improved |
| use DCAE-V instead of Wan2.2VAE under latent noise 0.2 | VAE reconstruction · Panda-70M 1000 samples | PSNR: 25.94 -> 29.34 | Improved |
| use DCAE-V instead of Wan2.2VAE without noise | VAE reconstruction · Panda-70M 192p | PSNR: 35.61 -> 33.25 | Hurt |
| quantize SANA-Video to NVFP4 via SVDQuant | SANA-Video · 720p 5s video on RTX 5090 | generation time: 71 s to 29 s | Improved |
| swap Wan2.1 VAE for DCAE-V (720p) | SANA-Video-2B · 720x1280x81 latency | latency: 36 vs 116 (Wan-2.2-5B) | Improved |
Later papers that compared with this method
4 later accepted papers report SANA-Video as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Attention Surgery: An Efficient Recipe to Linearize Your Video Diffusion Transformer (CVPR 2026 · better on 1 of 3 shared results)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · better on 1 of 2 shared results)
- ReHyAt: Recurrent Hybrid Attention for Video Diffusion Transformers (CVPR 2026 · better on 1 of 2 shared results)
- MobileWan: Closing the Quality Gap for Mobile Video Diffusion (NeurIPS 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SANA: Efficient High-Resolution Text-to-Image Synthesis with Linear Diffusion Transformers (ICLR 2025 · similarity 0.88)
- LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 2025 · similarity 0.88)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.88)
- Fast Video Generation with Sliding Tile Attention (ICML 2025 · similarity 0.88)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.87)
- SANA 1.5: Efficient Scaling of Training-Time and Inference-Time Compute in Linear Diffusion Transformer (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
