imagevolve

A variant is a program of two to six deterministic operations, such as tile shuffles, swirls or recolourings, each with its parameters, a strength and sometimes a soft mask. CLIP (opens in a new tab) ViT-B/32 scores it by how far its image embedding moves from the original's along a text direction: from the embedding of "a photograph" to that of your words, minus the mean of such directions over 430 generic phrases, so that changes every phrase rewards do not count. Nothing is optimized by gradient descent: CLIP only ranks the variants.

The search keeps the eight best variants that do not look alike, their image embeddings at a cosine below 0.92. Before you tap one, it proposes random programs, since automatic breeding collapsed onto one look in blind reviews. After a tap, three proposals in four mutate the tapped variant or its best refinement: a parameter changes, an operation is inserted, deleted or moved, or a few operations are spliced in from another variant.

On WebGPU, the operations run as shaders whose output goes straight to CLIP, with 4-bit weights and 160-pixel input, which scores about 1,100 variants a second on a $350 GPU. Your picture never leaves your device.