SANA: EFFICIENT HIGH-RESOLUTION IMAGE SYNTHESIS WITH LINEAR DIFFUSION TRANSFORMER

Summary from the TLDR AI newsletter

Sana is an efficient image generation model that can create high-quality 1024x1024 images in under a second on a laptop GPU. Its key innovations include a 32x image compression autoencoder (DC-AE), linear attention replacing vanilla attention in DiT, a decoder-only LLM for text encoding, and optimized training and sampling methods. The 0.6B parameter model matches or outperforms much larger models like Flux-12B while being 20x smaller and 100x faster. Sana-0.6B requires only 9GB VRAM for inference, making it accessible for consumer hardware. The repository includes code for training, inference, and evaluation, with both 0.6B and 1.6B model variants available.