mindview-t2i-turbo, in your browser

Type something and it paints it, in one, two or four steps, on your own graphics card. Nothing runs on a server: this Space is a static page, and the model, mindview-t2i-turbo, runs in your browser with WebGPU.

• You need a desktop browser with WebGPU (a recent Chrome or Edge) and a GPU with about 3 GB to spare.
• The first visit downloads the model (1.2 GB, one file). The browser keeps it for later visits.
• On a MacBook Air (M4) a first picture takes about 4 s in one step (Instant), 6 s in two (Fast) and 15 s in four (Best). Quick Best runs Best's first three steps at a quarter of the size, in about half the time.
• On a phone it would need more graphics memory (about 2.8 GB) than a phone gives a web page, so the page says so before it downloads anything.

The model is one file: the first 9 layers of PrismML's Ternary Bonsai 1.7B read the prompt, a linear map turns them into the painter's conditioning, the diffusion transformer of Bonsai Image 4B (FLUX.2 [klein] 4B with ternary weights) paints, and TAEF2 turns the result into pixels. Nearly every weight is −1, 0 or +1. Small side branches, trained on the ternary painter, let it paint in fewer steps. The model card has the details and the sources.

This is the paint page of mindview, an art installation that shows these models computing, layer by layer. The code is at github.com/NeoVand/mindview.