SANA-1.5

SANA-1.5 is NVIDIA Research's efficient text-to-image model (with MIT HAN Lab), announced January 29, 2025 with code and weights released that March. It replaces quadratic attention with a linear diffusion transformer plus a deep-compression autoencoder, so a 1.6B-parameter model generates 1024px images in about 0.1 seconds on an H100 (0.3s on an RTX 4090) — orders of magnitude faster than comparable diffusion models.

The codebase moved to Apache 2.0 in January 2025, and 4-bit quantized builds run in as little as 8GB VRAM. SANA-Sprint, a one/few-step distill, pushes latency even lower. Limitations: as a compact model it trails large diffusion systems on fine detail and complex compositions, and the training-data pipeline (WebDataset TARs) is heavier than the inference story.