So what is noise painting?
Noise painting is an adversarial input method that still works on most conventional diffusion models.
It starts with a hand-crafted piece of abstract art -- something designed with wild, unpredictable patterns and conflicting visual logic. The best ones mess with your eyes. They fool your brain into thinking there’s a signal hiding inside the chaos.
Then you feed it to a machine and wait for the magic to happen.
If you do it right, the denoiser malfunctions -- beautifully. The model starts hallucinating wildly, jumping to conclusions that don’t exist in the training data. It’s like arithmetic: there must be a sum for every input, even if that input points to a negative, unexpected address in latent space.
That's the first leg. Bending an AI around an unexpected unanticipated visual input.
Thing is, it’s not really fun unless you’re pairing it with an incongruent prompt that’s already stretching the model past what it should be doing. It's one thing to turn the denoiser sideways, but if you're not pushing the VAE to the edge of collapse... what’s the point?
Anyway. Taken together?
You get art.
You can also do it with music.
And now, you understand noise painting.
Whenever I test a fresh Stable Diffusion instance, I always run a few noise paints through to see if it's behaving.