apophenia
Random transformations of your picture, kept when they come closer to your words.
How it works
Apophenia is seeing meaning in random patterns, which is what CLIP does here. Each variant is a random program of up to eight image operations, such as tile shuffles, swirls or recolourings, some confined to a region. CLIP ViT-B/32, its vision model adapted to 160-pixel input, scores it by projecting the change in its image embedding onto a text direction, the embedding of your words minus that of "a photograph". The mean of such directions over 430 generic phrases is subtracted, so a change that every phrase would favour does not count. The page keeps the eight best variants that differ from each other, and tapping one makes the next ones its mutations.
Everything runs in your browser with ONNX Runtime on WebGPU or WebAssembly. The first visit downloads 103 MB, and your picture never leaves your device.