Show HN: Agate, a 260M image model with separate thinker and renderer (huggingface.co)

Agate, a new 260M-parameter text-to-image model developed by LogoLabs, has been unveiled as a potential game-changer for the AI/ML community, particularly in the realm of icon generation. This model, which took only 145 GPU-hours to train from scratch, achieves a competitive score of 0.550 on the GenEval benchmark, matching SDXL and significantly outperforming earlier models like SD 1.5 and 2.1. What sets Agate apart is its unique architecture that separates the image generation process into two components: a thinker (a small recurrent transformer) that plans the layout and a renderer (a convolutional U-Net) that paints the image. This division allows for precise object placement and color binding, making it especially adept at creating complex visual compositions. The significance of Agate lies in its efficiency and accessibility, able to run in under two seconds on a consumer GPU and even in the browser using WebGPU. This opens up opportunities for rapid design iterations and custom model training, making it easier for smaller teams and individual developers to create high-quality graphics without requiring expensive infrastructure. Additionally, Agate's MIT license facilitates further research and adaptation, allowing the community to build on its innovations without barriers. As it stands, Agate represents a promising advancement for smaller, specialized AI models, paving the way for more agile development in graphic design and related fields.