Overview
SANA-Video 2.0 is an efficient hybrid video diffusion Transformer designed for generating 720p high-quality videos on a single GPU. Through innovative hybrid linear attention and attention residual techniques, it achieves generation quality comparable to full softmax video DiT while maintaining the advantages of linear attention for long sequences. The model offers two scales, 5B and 14B, with a unified architecture balancing efficiency and performance.
Key Features
- Hybrid Linear Attention: Combines linear attention with attention residuals to avoid quadratic complexity, supporting efficient generation of long video sequences.
- Single GPU 720p Generation: Optimized design enables generation of high-quality 720p resolution videos on a single GPU, lowering hardware requirements.
- Unified Architecture with Multiple Scales: Offers 5B and 14B parameter versions sharing a unified architecture, flexibly adapting to different computing resources and quality needs.
Use Cases
- Video content creation and rapid prototyping
- Video generation experiments in research and education
- High-quality video generation in resource-constrained environments
Pros
- Runs on a single GPU, low deployment cost
- Generation quality comparable to full softmax models
- Efficient long-sequence processing, suitable for high-resolution videos
Pricing
Open-source model, free to use
Summary
SANA-Video 2.0 achieves efficient 720p video generation on a single GPU through hybrid linear attention and attention residual techniques, balancing quality and performance, making it suitable for high-quality video generation in resource-constrained scenarios.
Version History
- UEmbed：统一稀疏与稠密的多模态嵌入模型 (2026-08-03): UEmbed is a decoder-only multimodal embedding model that simultaneously generates sparse word-level and dense representations in a single causal forward pass, breaking through the single-token information bottleneck through learnable special tokens and vocabulary partitioning.