盆

Bonsai Image WebGPU

Text to image, 4 billion ternary parameters, in your browser.

ternary DiT · custom WGSL

Keep the model on this device

About 3 GB, downloaded once from PrismML's Hugging Face repository and kept in browser storage. Prompts and images never leave this device.

Checking WebGPU…

Fast encodes only the prompt's own tokens, as the community WebGPU demo does. Reference pads the prompt to 512 tokens exactly like PrismML's MLX and diffusers pipelines; it is slower and matches their images. Both use the model's 4 Euler steps.

—prompt
—4 denoising steps
—decode
—seconds per image
Device

Detecting adapter…