← sources

[RELEASE UPDATE] Agate-003-preview / better than SD 1.5 performance at 3x less parameters (<300M param including text encoder & VAE) with only 5000 EUR GPU cost.

source: reddit (reddit:StableDiffusion)published 2026-09-29 12:37ingested 2026-09-30 02:41

retained text

We've continued pretraining on the Agate-001-preview, releasing an intermediary Agate-002-preview (256x256) as well as a 512x512 native Agate-003-preview by continuing training while keeping the planning part at the same resolution as in 256x256 with a learned downsampler/upsampler.

https://preview.redd.it/9s48sue2egsh1.png?width=1080&format=png&auto=webp&s=e900381de31433c564f913438d422f0ed2a383cf

The model is SoTA in it's parameter count area, using two approaches which we credit for the improvements observed:

1. A planner architecture, that plans in low resolution before submitting the task to the "painter / renderer". The planner is a transformer and cross-attends to the text tokens, while the renderer is a fully convolutional approach which comes with very good image priors.
2. A weighted loss that uses P-Dino / LPIPS measurements and saliency for weighing to relevant areas (people look more at faces / text so these must be prioritized in terms of loss vs the background).

https://preview.redd.it/4r9etzmhegsh1.png?width=1080&format=png&auto=webp&s=c223f05f7e0f871f19e7c7aa49f3f0127a8d120e

https://preview.redd.it/o74odz4iegsh1.png?width=1080&format=png&auto=webp&s=f319a90e7ddc588cd38bad41c399ee1cdd904206

The whole training run + R&D to test the architecture took under 5000 EUR, with an expenditure of only 630 GH-200 GPU hours.

https://preview.redd.it/7dxytd8negsh1.png?width=2560&format=png&auto=webp&s=f03bd94c8ffe9c2eb4efa61f95ca41d3bc96d74d

The training lineage itself cost slightly over half, at 319 GPU-h.

The main focus was testing this architecture for efficient generation of images, and we plan to use it further for font / glyph / symbol generation as part of our work on Logolabs.

The model, a generalist, supports LoRA training that can significantly improve quality on specific tasks and design styles. With the model being teeny, LoRA training is very accessible.

https://preview.redd.it/0ohvpekvegsh1.png?width=1080&format=png&auto=webp&s=aec5a2920d0c3c5ff399e1f50d9e87f232f04d5e

https://preview.redd.it/dcbxqe2wegsh1.png?width=1080&format=png&auto=webp&s=3c4170e3dbf9ff6c838e32babfe16dafd682ea76

https://preview.redd.it/q6npo8hwegsh1.png?width=962&format=png&auto=webp&s=755d57fe6e51f865f299ababa09afe21bba39a5d

Technical details on HuggingFace: [https://huggingface.co/Logolabs/agate-preview-003](https://huggingface.co/Logolabs/agate-preview-003)

Try on WebGPU: [https://huggingface.co/spaces/Logolabs/agate-webgpu](https://huggingface.co/spaces/Logolabs/agate-webgpu) (including thinker preview)

ComfyUI Nodes: [https://github.com/logolabs/agate-comfyui](https://github.com/logolabs/agate-comfyui)

---

It is mainly intended to be used after being LoRA finetuned, for narrow tasks that require very low latency or need to work even on CPUs.

It's entire pipeline is sub 300M param, while other small models offload compute by having a massive text encoder.

If you want to support efforts like ours, we would be glad if you can come with constructive criticism and/or a star on Hugging Face!

Thank you!

document fetched 2026-09-30 00:35 · canonical https://reddit.com/r/StableDiffusion/comments/1wt8z9j/release_update_agate003preview_better_than_sd_15