Source-linked AI summary
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, Hao Zhang, Muyang Li, Yukang Chen, Han Cai, Sanja Fidler, Ping Luo, Song Han, Enze Xie
TL;DR
High-resolution and long video generation are limited by token growth, attention cost, memory use, and expensive training. SANA-Video addresses these constraints with Linear DiT, constant-memory block attention, and efficient data and training strategies, achieving competitive quality with substantially faster and lower-cost generation.
Problem
High-resolution and long video generation process many tokens, making O(N^2) self-attention, memory use, and full-sequence processing major computational bottlenecks.
Method
SANA-Video combines linear attention in a video diffusion transformer with block-wise autoregressive generation using a fixed-memory KV state, pretrained initialization, data filtering, and staged training.
Results
SANA-Video delivers competitive benchmark performance while being 16× faster than modern small diffusion models and requiring 12 days on 64 H100 GPUs for training.
Takeaways & Limitations
SANA-Video provides a low-cost, high-quality foundation for fast high-resolution and minute-long video generation, including deployment on RTX 5090 GPUs.
Abstract
from arXiv · showhide
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.
1. Introduction
SANA-Video targets the computational and memory costs that make high-resolution and long video generation difficult. It combines linear attention, block-wise constant-memory caching, and efficient training practices to deliver faster, lower-cost generation.
- Motivation: Wan 14B requires over 75,000 tokens and 32 minutes on an H100 to generate a single 5-second 720p video.Large token volumes and full-sequence processing also make videos longer than 10 seconds difficult to generate efficiently.
- Efficiency: 12 days on 64 H100 GPUs represent as little as 1% of MovieGen’s training cost, while 5-second 720p inference takes 36 seconds on an H100.The paper presents this efficiency as making high-quality video generation more accessible and practical.
- Linear DiT: Linear DiT replaces O(N^2) vanilla self-attention with O(N) linear attention, producing a 4× acceleration on 720p video.The video design adds RoPE for long-context modeling and a 1D temporal convolution in Mix-FFN.
- Long-video generation: Block Linear Attention reduces the KV cache to small, fixed memory and fixed computation per new token for long-context generation.The design supports a two-stage autoregressive training paradigm for long videos.
- Training: SANA-Video’s low training cost relies on a pretrained SANA T2I model, source-specific data filters, detailed video captions, and an efficient training strategy.These choices are presented as the main contributors to the model’s reduced training cost.
- Results: SANA-Video is over 13× faster than Wan2.1 for 720p generation while achieving competitive results across many benchmarks.With NVFP4 on an RTX 5090, it generates a 5-second 720p video in 29 seconds.
2. Preliminaries
Autoregressive diffusion models combine block-wise chain-rule decomposition with denoising diffusion to model sequential dependencies in long video generation. Each block is generated conditionally on preceding blocks.
- Autoregressive diffusion: Autoregressive diffusion models combine token- or block-wise autoregressive decomposition with denoising diffusion models.This framework is used for long-sequence generation in domains including language and video.
- Chain-rule formulation: For a sequence of N blocks, generation factorizes into conditional block distributions using the chain rule.Each block distribution is modeled with a diffusion process.
- Rationale: The approach combines sequential-dependency modeling from autoregression with block-wise high-quality generation from diffusion.The passage describes these as complementary strengths of the framework.
3. SANA-Video
SANA-Video adapts linear diffusion transformers and block-wise causal processing to make high-resolution and long-video generation more efficient. Its training, attention, and compression designs target lower computational and memory costs while retaining global context and enabling fast generation.
- Linear DiT: O(N^2) self-attention becomes a bottleneck as video resolution and sequence length grow, motivating linear attention with O(N) complexity.Linear Video DiT extends SANA’s architecture with RoPE and temporal convolution for video modeling.
- Linear DiT: RoPE is applied after the ReLU kernel to improve positional modeling while producing sparser, more localized attention patterns.The design prevents the ReLU kernel from filtering out positional information.
- Linear DiT: Removing RoPE from the denominator keeps the QK-sum denominator positive and stabilizes training while retaining positional encoding in the numerator.The numerator includes RoPE on queries and keys, whereas RoPE is removed from either the key or query in the denominator.
- Block Linear Attention with KV Cache: Causal linear attention compresses prior-token information into a constant-size state, providing global memory with O(D^2) memory and computation.Because N > W >> D, causal linear attention is described as more efficient than causal full and local attention while maintaining global memory.
- Block Linear Attention with KV Cache: Block causal Mix-FFN uses zero padding and causal temporal convolution to prevent future-block leakage and aggregate information from adjacent blocks causally.The temporal convolution uses kernel size 3 and requires the last frame of the preceding block.
- Efficiency and Results: 27 FPS generation speed enables 1-minute, 16 FPS 480P videos within 35 seconds on an NVIDIA H100 GPU using 4-step LongSANA.The same efficiency strategy also supports 720P generation, including 36 seconds for a 5-second video and 53× acceleration over Wan2.1-14B.
4. Experiments
Experiments show that SANA-Video combines competitive video quality with strong efficiency, while architectural choices improve attention performance, high-resolution latency, and block consistency.
- Performance Comparison and Analysis: 83.71 Total Score on text-to-video and 88.02 Total Score on image-to-video demonstrate competitive VBench performance.The image-to-video score outperforms Wan2.1 (14B) and HunyuanVideo-I2V (11B), while text-to-video is comparable with Open-Sora-2.0 (14B) and exceeds Wan2.1 (1.3B).
- Performance Comparison and Analysis: 60 seconds latency for a 480P 81-frame video makes SANA-Video the fastest model in the reported comparison.This corresponds to 7.2× higher throughput than MAGI-1 and over 4× higher throughput than Step-Video.
- Ablation Studies: 3D RoPE lowers training loss, while temporal convolution in Mix-FFN significantly enhances training performance.These architectural modifications are evaluated through training-loss curves in Fig. 6(a) and Fig. 6(b).
- Ablation Studies: 2× speedup at 480P and 4× at 720P show that linear attention becomes increasingly advantageous at higher resolutions.The comparison is between SANA-Video’s linear and full-attention configurations.
- Ablation Studies: Monotonically increasing SNR sampling produces better quality and more consistency across autoregressive blocks than random timestep sampling.The comparison uses different blocks in Fig. 6(d).
- Long Video Generation: SANA-Video matches Self-Forcing and outperforms SkyReel-V2 and CausVid on long-video VBench evaluation.The comparison is reported for autoregressive video generation methods.
5. Applications and Deployment
SANA-Video serves as a pre-training model for several video-generation applications and is adapted for efficient edge inference through NVFP4 quantization.
- Applications: SANA-Video is extended to embodied AI, autonomous driving, and game-generation world-model applications.These applications are presented as adaptations of the pre-training model.
- Deployment: Figure 7 compares SANA-Video latency under BF16 and NVFP4 precision.The figure focuses on precision-dependent inference latency.
- Deployment: NVFP4 quantization is used to support efficient inference and on-device deployment.The quantization selectively targets projection and feed-forward convolution layers while retaining other components in higher precision.
- Deployment: The deployment design balances efficiency and fidelity through selective rather than universal quantization.Normalization layers, temporal convolutions, and cross-attention KV projections are kept unquantized.
6. Related Work
Related work frames video generation as diffusion-based latent modeling and surveys efforts to add temporal structure, autoregressive planning, and efficient linear-complexity attention.
- Video Diffusion: Early video-generation studies adapted text-to-image models by adding temporal layers, while later methods advanced latent video diffusion.The cited examples include Make-A-Video, PYoCo, Tune-A-Video, MagicVideo, SVD, and Latent Video Diffusion.
- Autoregressive Diffusion: Recent research combines autoregressive decomposition with diffusion to pursue long-term planning for visual generation.Autoregressive diffusion models apply token- or block-wise chain-rule structure alongside denoising diffusion.
- Efficient Attention: DiT has become a mainstream architecture, while linear-complexity mechanisms address vanilla attention’s O(n^2) computational cost.The discussion includes PixArt-α, Stable Diffusion 3, Flux, and gated linear attention methods.
7. Conclusion
The paper concludes that SANA-Video enables high-resolution, high-quality, long-video generation with low training cost, fixed-memory long-context support, and substantially faster speed.
- Conclusion: SANA-Video uses linear attention, block linear attention with a constant-memory KV cache, and efficient data and training strategies.Together, these designs target token-extensive video generation, minute-long sequences, and reduced training cost.
- Conclusion: 12 days on 64 H100 GPUs narrows training cost while maintaining competitive performance with modern small diffusion models.The conclusion reports this as a central efficiency-performance result.
- Conclusion: SANA-Video is reported as 16× faster than modern state-of-the-art small diffusion models while retaining competitive performance.The claimed comparison is made at the model level rather than for a single benchmark entry.
A. LLM Usage
LLMs were used only for editorial assistance in preparing the manuscript, not for its scientific content or analysis.
- LLMs refined grammar and phrasing to improve the manuscript’s clarity and readability.
- LLMs helped improve logical flow between sections and condense overly verbose passages.
- The authors conceived and executed all research ideas, experimental designs, and data analyses without LLM contribution.
B. More Implementation Details
SANA-Video modifies the SANA architecture for video generation and evaluates VAE robustness under latent noise perturbations.
- Architecture: SANA-Video increases the FFN and attention head dimensions to support 3D RoPE and adds temporal convolution for motion performance.
- Architecture: The architecture details are summarized in Table 6.
- VAE Evaluation: The paper analyzes several video VAEs by testing reconstruction after adding Gaussian noise to encoded latents.
- VAE Evaluation: DCAE-V performs more robustly across noise levels, indicating stronger reconstruction generalization for the diffusion model.
C.2. Qualitative Comparison
Qualitative results show that SANA-Video provides strong semantic alignment and motion control across text-to-video and image-to-video generation, including long videos.
- Text-to-Video Generation: SANA-Video has comparable semantic understanding, motion control, and aesthetic quality to small text-to-video diffusion models.
- Image-to-Video Generation: SANA-Video shows the best semantic understanding and motion control among the compared image-to-video models.
- Unified Generation: The unified model generates semantically consistent and temporally smooth videos conditioned on a reference image’s first frame.
- Motion Control: Increasing the motion score produces larger but still consistent motion in image-to-video generation.
- Long Video Generation: LongSANA generates one-minute videos with motion consistency and semantic alignment.
D. Data Processing Pipeline
The data pipeline combines scene segmentation, quality filtering, caption rewriting, and curated SFT selection to prepare training videos.
- Pipeline Overview: The pipeline creates single-scene 5-second clips and filters data using motion, aesthetic, and saturation criteria.
- Scene Detection and Shot Cut: PySceneDetect and FFmpeg split raw long videos into scenes and short clips for pre-training.
- Motion Filtering: Unimatch and VMAF identify videos with moderate and clear motion, while motion scores are appended to prompts for control.
- Aesthetic and Saturation Filtering: DOVER provides the overall aesthetic score used for filtering, while OpenCV measures saturation to remove unnatural-color videos.
- Captioning and SFT Data: Qwen-2.5-VL rewrites captions, and approximately 5,000 high-quality videos are selected for final SFT training.
E. World Model
SANA-Video is fine-tuned for world-model-related generation across embodied AI, autonomous driving, and game generation. The paper presents simulation and generation applications using robotics, driving, and Minecraft data.
- Downstream Tasks: SANA-Video is fine-tuned for embodied AI, autonomous driving, and game-generation tasks.These downstream tasks demonstrate its potential for world-model-related generation.
- World Model for Embodied AI: For embodied AI, SANA-Video generates simulation data for robot training from synchronized multi-camera data.The head-front view is used as the target video after filtering.
- Visualizations: The paper includes visualizations of image-to-video consistency, motion control, long-video generation, captioning, SFT effects, and world-model outputs.The figures cover realistic motion, motion-score effects, LongSANA, captioning, human-preferred fine-tuning, and downstream generations.
- World Model for Autonomous Driving: For autonomous driving, SANA-Video generates diverse and realistic driving scenes using front-camera driving data.The fine-tuning data uses a front camera with 30 FOV.
- World Model for Game Generation: For game generation, SANA-Video is trained on Minecraft player recordings to create interactive video games.The raw videos are processed through the data pipeline, and a classifier identifies low-quality scenario data.