Demo of the week

by webml-community

Bonsai Image WebGPU runs image generation entirely in the browser using WebGPU, drawing on 1-bit and 2-bit quantized 4B-parameter models to keep compute requirements low enough for client-side inference. The quantization approach is worth examining closely if you are exploring how far model compression can go before output quality degrades. Builders working on edge or offline deployment pipelines can use this as a reference point for what ternary and binary weight schemes look like in a real, running system.