SANA-Video: Efficient Video Generation with Block Linear Diffusion Transformer
Paper · https://arxiv.org/abs/2509.24695
We introduce SANA-Video, a small diffusion model that can efficiently generate videos up to 720x1280 resolution and minute-length duration. SANA-Video synthesizes high-resolution, high-quality and long videos with strong text-video alignment at a remarkably fast speed, deployable on RTX 5090 GPU. Two core designs ensure our efficient, effective and long video generation: (1) Linear DiT: We leverage linear attention as the core operation, which is more efficient than vanilla attention given the large number of tokens processed in video generation. (2) Constant-Memory KV cache for Block Linear Attention: we design block-wise autoregressive approach for long video generation by employing a constant-memory state, derived from the cumulative properties of linear attention. This KV cache provides the Linear DiT with global context at a fixed memory cost, eliminating the need for a traditional KV cache and enabling efficient, minute-long video generation. In addition, we explore effective data filters and model training strategies, narrowing the training cost to 12 days on 64 H100 GPUs, which is only 1% of the cost of MovieGen. Given its low cost, SANA-Video achieves competitive performance compared to modern state-of-the-art small diffusion models (e.g., Wan 2.1-1.3B and SkyReel-V2-1.3B) while being 16x faster in measured latency. Moreover, SANA-Video can be deployed on RTX 5090 GPUs with NVFP4 precision, accelerating the inference speed of generating a 5-second 720p video from 71s to 29s (2.4x speedup). In summary, SANA-Video enables low-cost, high-quality video generation.
First seen by Oriel on .
AI summary
SANA-Video is a small diffusion model for generating high-resolution videos up to minute-long duration, using linear attention and constant-memory state to reduce inference and training costs.
- Its Linear DiT uses linear attention to process the large token counts involved in video generation more efficiently than vanilla attention.
- A block-wise autoregressive method maintains global context with a constant-memory state, replacing the traditional KV cache for long video generation.
- Training took 12 days on 64 H100 GPUs, which the authors report is 1% of MovieGen's cost.
- SANA-Video has competitive performance against Wan 2.1-1.3B and SkyReel-V2-1.3B while achieving 16x faster measured latency.
- On RTX 5090 with NVFP4 precision, generating a 5-second 720p video takes 29 seconds instead of 71 seconds, a 2.4x speedup.
#video generation #diffusion models #linear attention #kv cache
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hugging Face Papers: 56, latest observation ; 0 over observed period since
Comments
Comment count reported by the source.
- Hugging Face Papers: 2, latest observation ; 0 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Papers | 2509.24695 | https://arxiv.org/abs/2509.24695 Read discussion |