How SANA 1.5’s Efficient Linear Diffusion Transformer Sets a New SOTA in Text‑to‑Image Generation

The paper introduces SANA 1.5, an efficient linear diffusion transformer that scales training and inference compute via model growth, depth‑wise pruning, and inference‑time scaling, achieving a GenEval score of 0.80 and matching larger models while using far less resources.

AISANAdiffusion

0 likes · 23 min read