No.150 · AI & Chat
Diffusion Preview
Image generation shown as it happens: four candidates start as coloured noise and resolve step by step, with the step count and seed alongside.
The prompt
The design brief behind this component — reuse it to regenerate, remix, or explain it.
Build an image-generation card that shows denoising. A prompt row with a Generate again button; below, a large square canvas and a side column with four candidate thumbnails (a radiogroup), step/seed/sampler readouts and a thin progress bar. Each candidate has a target picture (here a generated canvas study per seed) and two noise textures from its seed — a 12×12 colour noise drawn up smoothly (the blotchy latent look) and a 96×96 noise drawn crisp (grain). Each step draws the target with a blur that shrinks from 22px to 0 and saturation rising, then the coarse noise fading out, then the fine grain fading out, with a = k^1.1 (coarse noise at up to 72%, grain at up to 36%) so shapes appear before detail. Thirty steps at 85ms; each candidate runs 1.5 steps behind the previous, so they resolve in a ripple. The large canvas mirrors the picked candidate; picking another after generation shows it large. Generate again bumps every seed and starts over; under reduced motion the finished images appear at once.
Notes
- Interaction
- Watch the four candidates resolve; pick one to see it large; Generate again for new seeds.
- Motion
- 30 steps at 85ms; candidates staggered by 1.5 steps; blur 22px → 0 and noise fading on k^1.1.
- Accessibility
- Candidates are a labelled radiogroup (with seeds in their names); the step count is announced politely and the large canvas is role="img" with the prompt. Reduced motion shows finished images.
- Responsive
- A card up to about 40rem wide that fills narrower screens; dense rows and side columns stack on phones.
- On touch
- Everything is tap-driven; hover hints have tap or focus equivalents.