WebGPU diffusion compare — two models (browser ONNX). v 1.0 Jonathan Rothberg 2026

No Python server: inference runs in your browser via web-txt2img (workers + ONNX Runtime Web / WebGPU). Models: fast single-step sd-turbo, janus-pro-1b (Transformers.js ONNX on WebGPU), and sd-1.5 (25-step, higher quality, ~2GB).
Serve this file over http(s) (not file://) so workers, WASM, and Hub downloads work. First run downloads ~2.3GB per model (cached).
VRAM: each column uses a separate worker; loading both models at once may exceed GPU memory — unload one side or use one model at a time.

Model A

Model B

Prompt

Seed (SD-Turbo & SD 1.5; Janus ignores)

Steps (SD 1.5 only)

Guidance scale (SD 1.5)

WebGPU diffusion compare — two models (browser ONNX). v 1.0 Jonathan Rothberg 2026

Model A output

Model B output