Ready — 640 × 400 px
Whatever runs hands back a matte — a coverage map saying how much of each pixel is subject. Every setting afterwards reshapes that matte on the spot, so nothing has to be worked out twice.
Starting…
Google's 2015 DeepDream, run on your own machine. The picture is pushed towards whatever one layer of an image-recognition network already thinks it can see, over and over, until the network's own imagination grows out of it — dogs, birds and eyes, because that is what its training pictures were full of.
The first run fetches the network (about 12 MB) from this site; your browser then caches it. Nothing is uploaded.
Starting…
The layer shows each octave as it finishes. Cancel puts the picture back.
Model: inception5h · Apache-2.0 · Engine: TensorFlow.js
DeepSeek's Janus-Pro, drawing on your own machine. It writes a picture the way a language model writes a sentence — one token at a time, 576 of them — and turns the finished sequence into an image at the end, which is why the wait is a count rather than a picture slowly clearing.
Starting…
Nothing is touched until you press Apply.
The seed makes a picture repeatable: the same words, seed and settings give the same picture again, and asking for several numbers them upwards from it. Variety is the sampler's temperature — low sticks to the obvious reading of the words and repeats itself, high wanders; 0.7 is what the model ships with. Choices is how many of the likeliest tokens are in the running at each step, out of a picture vocabulary of 16,384; the runtime's default of 50 is a narrow field, and opening it up loosens the picture.
Model: Janus-Pro-1B · DeepSeek Model Licence · Engine: Transformers.js
Paste an array to see it.
Any file will do — a picture, a font, a recording, a program. Drop one here.
Choose a file to see it.