Sana

Why we included this project

Sana is NVIDIA's text-to-image project built around a linear diffusion transformer, a design that keeps image quality high while cutting memory and compute well below typical diffusion models. A 0.6B model can generate detailed images on a single consumer GPU, and the 4-bit variant runs in under 8GB, so a small team can actually run and fine-tune it without renting a cluster of H100s. The repository covers both sides of the work: training and inference guides for getting models up and running, plus companion models for controllable generation, video, streaming, and reinforcement-learning alignment. For teams that want production-grade image synthesis on hardware they already own, this is a practical place to start.