Sana

Sana generates text-to-image output at up to 4096x4096 using a linear-attention diffusion transformer from NVIDIA research that trades quadratic self-attention for linear complexity at high resolution. The family spans Sana 0.6B and 1.6B, SANA-1.5 at 1.6B and 4.8B parameters, SANA-Sprint for one-step or few-step generation at around 0.1 seconds per 1024px image on an H100, and SANA-Video at 2B for 720p clips. The repository ships training and inference code, LoRA and DreamBooth fine-tuning, a ControlNet implementation, FSDP distributed training, and integrations with Diffusers, ComfyUI, and SGLang. NVIDIA reports 4K generation inside 22 GB of GPU memory with model offload, while 4-bit quantization brings inference within 8 GB of VRAM. Code is Apache-2.0, but pretrained weights on Hugging Face carry NVIDIA license terms oriented toward research use, so commercial deployment needs a close read of the specific model card. The paper was accepted as an oral at ICLR 2025 and the repo sits around 8.6k stars, making it a reference point for efficient high-resolution image synthesis.