[NEW RELEASE] Agate-001-preview, 260M parameter (including text encoder), pareto frontier model with close to SD 1.5 performance. MIT licensed & open weight. Novel thinker-renderer architecture.
retained text
https://preview.redd.it/hms6dpiyhprh1.png?width=1672&format=png&auto=webp&s=5b71cab7e350ed02d3824566e76d28ae412e8734 We're launching AGATE-001-PREVIEW, our attempt at maximizing the parameter-efficiency of current AI models. Our approach has yielded pareto-frontier results, with the model significantly outperforming models twice it's size, with the closemost competitor being SD 1.5 with 3x more parameters in total. https://preview.redd.it/kqkanmw8iprh1.png?width=2912&format=png&auto=webp&s=21bc2eb9e6d4f43ee791caac8f7d7532136646b6 While the model is limited to 256x256 px, and uses an older VAE (SD-1.5 VAE), it is showing impresive learning capacity, being able to mix complex concepts, understand spatial arrangement, and more. https://preview.redd.it/opm13luciprh1.png?width=1650&format=png&auto=webp&s=80157dd1e804ea4f602fe56ba0ecf78f02732f09 The model was trained on Flux-Reason-6M, for \~26 epochs, and has not converged, **meaning this is a release preview and meaningful imrpovements in performance could come from continuous training**. We've managed to significantly surpass DiT and FDCM internally by splitting the model in two stages, one for general planning, one for drawing. Transformer models are great at planning and overall understanding of a canvas, while convolutional solutions like FDCM are great at drawing. We use this to build a hybrid architecture, where an FDCM renderer is responsible with converting latent thinking tokens into images. The thinker is a small recurrent transformer that reads the prompt and plans a 16 × 16 region map. The thinker cross-attends to the text encoder, that is trained jointly after a warmup of 10 epochs where it is kept frozen. We've seen significant improvements in loss reduction after unfreezing the tiny text encoder, based on Ettin-68M. **The architecture is described below:** https://preview.redd.it/vsj1i0bziprh1.png?width=3000&format=png&auto=webp&s=a4e5bcc345695782ce7bd7766fbfd0f3c4d9d977 We've seen pareto frontier results at this size, with the model being quite competitive. https://preview.redd.it/rxqw5f5ajprh1.png?width=2380&format=png&auto=webp&s=83ae24ad0b07ee1e435f6962c3a1924df0a2e24f While the model is still beaten by larger models, it holds it's own against models 3x it's size overall. https://preview.redd.it/akmynb28jprh1.png?width=2360&format=png&auto=webp&s=84f9fd4f83912d8615039a39c60f9bf0dec4e993 Below, you can see some generations: https://preview.redd.it/arkw4hrdjprh1.png?width=2014&format=png&auto=webp&s=2287a041eac9a8670b2d356a0910350058b3d45a https://preview.redd.it/1s3ugt9ejprh1.png?width=2014&format=png&auto=webp&s=67f2cfb0baa148d9ee999bcf90b1e7c4b7c5362a Additionally, you can see architectural differences between various architectures we've used off-the-shelf, or adapted: https://preview.redd.it/3b59j9oijprh1.png?width=2010&format=png&auto=webp&s=f29444f84f74ae3ec3764908cea1c01f7b818fe5 **Few utilities of this solution:** * **Creating synthetic data that is cheap**, after being finetuned on a specific dataset. More so, the architecture can easily be adapted to do generative denoising / editing to help create massive synthetic data for classifier models, that would otherwise be very expensive to build with big models. * **Training small models on proprietary datasets**, as our solution is stronger per-parameter than the verified other architectures (DiT, FCDM). * Thumbnail generation (when blurred) for searching / loaders or similar low-strength solutions. * **On-Mobile generation.** **The model is open weight, and can be tried for free:** * Weights: [https://huggingface.co/Logolabs/agate-preview-001](https://huggingface.co/Logolabs/agate-preview-001) * WebGPU (try in your browser): [https://huggingface.co/spaces/Logolabs/agate-webgpu](https://huggingface.co/spaces/Logolabs/agate-webgpu) * ComfyUI nodes: [https://github.com/logolabs/agate-comfyui](https://github.com/logolabs/agate-comfyui) (Also allow peering into the thinkers' thinking tokens) If you want to help us out, it would be amazing if you can leave a heart on our huggingface repository and/or our WebGPU. **EDIT: WE got a ComfyUI node set for this!**
document fetched 2026-09-26 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1wq3oir/new_release_agate001preview_260m_parameter