Logolabs' Agate: 260M-Param Text-to-Image Model Trained in Just 145 GPU-Hours, MIT Licensed
linoy_tsaban · x · 2026-09-29
- Logolabs released Agate Preview, a text-to-image model with a split architecture: a transformer "thinker" plans a 16×16 region map from the prompt while a U-Net "renderer" paints with convolutions.
- Only 260M parameters and trained from scratch in 145 GPU-hours; runs in under 2 seconds on a consumer GPU.
- MIT licensed, with a WebGPU demo and open weights available.
Related event: Logolabs Unveils Agate, a 260M-Parameter Text-to-Image Model Rivaling SD1.5(2 posts)→
More from Multimodal
- Prompt Template: Duotone Cosmic Vision for Striking Two-Color Space Imagery — LudovicCreator · 2026-09-29
- One reference, a hundred styles: Recraft V4 Style demo with prompts — OVolosin82152 · 2026-09-29
- Kling 4.0 full version shows off native 30s clips, cinematic 21:9 in short film — hq4ai · 2026-09-29
- 300M-param image model trained for €5000 claims SD 1.5-level performance — incorporo · 2026-09-29
- Kling 4.0's long-sequence consistency makes it feel like a real filmmaking tool — SimplyAnnisa · 2026-09-29
- Finegrain Agentic Camera Demos Blazing Speed vs ChatGPT Image Processing — mattturck · 2026-09-29