Agate Preview 001

Agate Preview 001 is a lightweight text-to-image model that generates 256×256 images from English prompts. Designed and trained by LogoLabs, it features a unique architecture splitting generation into planning (transformer-based thinker) and rendering (convolutional U-Net renderer), enabling fast inference on consumer GPUs and in-browser use. It is MIT licensed, supports retraining and fine-tuning on smaller budgets, and focuses on precise object placement and color binding, ideal for icon and logo design workflows.