Step from pure noise to final image and watch structure emerge one denoising pass at a time. The viewer makes iterative denoising inspectable — slowing down the process so you can name what the prompt is guiding and what the model is deciding on its own.
Live preview · launch for the interactive version
Images don’t appear all at once. Each step removes a small amount of noise, pushing pixels toward a prompt-consistent distribution — one pass at a time.
Composition appears before subject, subject before texture, texture before detail. The sequence reveals what the model is building toward — and in what order.
The prompt doesn’t just shape the final output. It guides every denoising step. You can watch it pull structure out of noise — or watch where it stops mattering.
Don’t skip to the end. Pause at each stage and describe only what you can see — not what you expect.
Write down which step you expect the subject to become recognizable. Then step through and see if you were right.
At each stage: what regions exist? What subject? What texture? What detail? Describe only what you can observe, not what you infer.
Run the same sequence with a vague and a specific prompt. At which step does the difference become visible?
What did the model generate that your prompt left open? That unspecified choice is a default — name it precisely.