260M-parameter open-source image model Agate nears SD 1.5 quality, MIT licensed
incorporo · reddit · 2026-09-26
Logolabs released Agate-001-preview, a 260M-parameter (including text encoder) open-weight image generation model under MIT license that matches models several times its size, with SD 1.5 as the closest competitor. It uses a novel thinker-renderer architecture: a small recurrent transformer plans a 16×16 region map from the prompt, while a convolutional FDCM renderer converts latent thinking tokens into images, outperforming pure DiT and FDCM at this scale. Trained 26 epochs on Flux-Reason-6M (not yet converged), it is limited to 256×256 with the SD 1.5 VAE. Target uses include cheap synthetic data generation, small models on proprietary datasets, thumbnails, and on-mobile generation. Weights, a WebGPU browser demo, and ComfyUI nodes are available.
Related event: 260M-Parameter Open Model Agate-001 Nears SD 1.5 Performance(2 posts)→
More from Multimodal
- Midjourney SREF pick: --2523103199 for loose pencil-sketch portraits — tisch_eins · 2026-11-11
- Gemini-generated endless loop of quantum cryostats self-assembling — hapticdata · 2026-11-11
- Dad brings his daughter's crayon drawings to life with AI for $1.75, and it's unsettling — Illustrious_Ad_4292 · 2026-11-11
- MOSS-Image-SFT vs MiniMax-H3-Image (Pruned): 1:1 image benchmark across 10 art mediums — SkirtSwimming8950 · 2026-11-11
- UniPath (NeurIPS 2026): adaptively coordinating understanding and generation in one model — jindong_wang92 · 2026-11-11
- An AI-built RPG where a PhD student battles peer reviewers — s_scardapane · 2026-11-11