Agate 4-step WebGPU
A 0.19B text-to-image model distilled to 4 steps with no guidance pass, running on your own GPU. Nothing leaves your machine. Turn on Live and the image redraws as you type.
This page needs WebGPU (recent Chrome or Edge). You can still run it on the CPU, which takes a few seconds per image.
Not loaded
AI-generated image. It carries an invisible watermark and provenance fields in the saved PNG.
What this is
ML-Intern-lab/agate-preview-002-4step is an unofficial distillation of Agate Preview 002 by LogoLabs. Classifier-free guidance was baked into the weights, then the 50-step sampler was halved four times (32 → 16 → 8 → 4). Each image is 4 network passes instead of Agate's 100. On the official GenEval scorer (4 images per prompt) it gets 0.536, against 0.563 for the teacher at 50 steps; the teacher itself run at 4 steps gets 0.466.
256 × 256 only. No negative prompts (guidance is baked in). It inherits Agate's weaknesses: exact text, counts above three, negation. There is no safety filter in this page, and the base model can draw people partly unclothed without being asked.
Runtime adapted from LogoLabs' agate-webgpu (MIT): tokenizer, text encoder, sampler and TAESD decoder run through onnxruntime-web. Seeds use a JavaScript PRNG and do not match PyTorch seeds. This page is not affiliated with LogoLabs. Licence: MIT.