Sana

Sana is an open-source text-to-image foundation model developed by NVIDIA that focuses on efficient, high-quality image generation. Built with a rectified flow transformer architecture, it is designed to produce detailed, photorealistic, and stylistically diverse images from natural language prompts while maintaining strong training and inference efficiency. Sana supports multiple resolutions, including high-resolution outputs, and is optimized for modern GPU hardware, making it suitable for both research and production environments.

Key capabilities include precise prompt adherence, fine-grained control over visual attributes, and robust performance across a wide range of concepts, from everyday scenes and objects to complex compositions and artistic styles. The model is released with reproducible training recipes, reference implementations, and configuration details, enabling researchers and engineers to study, adapt, and extend the architecture. Sana also emphasizes scalable training, offering insights into data pipelines, optimization strategies, and distributed training setups.