Image reveal

Tab

Moves to the Retry button in the error state. The canvas itself is not focusable.

EnterorSpace

Activates Retry.

• The canvas has role img; its label names the state while generating and becomes the alt text once done.
• A polite live region announces Queued, Generating image, Image ready and the error message, not every step. Auto mode appends simulated.
• While generating, the pill is a progressbar with aria-valuenow and an aria-valuetext of Step n of total.
• Status is never color alone: the pill pairs a ring or check with a short label, and the error state pairs an alert icon with its message.

• Queued is the calm surface tone with one broad, soft shimmer that starts and ends fully off canvas, so its 2.6s loop has no seam.
• Generating starts from pure latent noise: per-cell gaussian color at about 1/8 of the output resolution, drawn as crisp blocks, with a new seeded frame every step crossfaded in 35ms.
• Under the noise is the model's current prediction, built from the image itself (6 to 96 px wide copies, smoothly upscaled). It moves along that chain per step, so color blobs come first, then composition, then shapes.
• Noise strength follows a schedule that drops fast early, so the composition reads by about a third of the steps and the last steps only refine. Steps run at a constant rate, like a real sampler.
• After the last step a tiled decode lands the sharp image tile by tile in raster order, each tile crossfading in, and the pill switches to Decoding.
• One soft-light sheen sweeps across once at completion (1.1s), the pill shows the elapsed time, then fades out.
• Starting over or swapping the photo fades the decoded image straight back to the surface before anything resets underneath, and a photo swap crossfades from a snapshot, so there is never a hard cut.
• Pill labels crossfade (100ms out, 160ms in after a 60ms gap) and the pill width springs with no bounce.
• Reduced motion keeps one static noise frame, decodes as a single crossfade and drops the shimmer travel, spinner and sheen.
• The frame loop stops off screen (IntersectionObserver) and in background tabs, and draws nothing once the image is resolved.

• Fills its container width (up to 560px by default) and keeps the given aspect ratio.
• The canvas renders at device pixel ratio (capped at 2), so it stays crisp from a 120px thumbnail to a large canvas.
• Latent noise is sized in CSS pixels (about 6 px cells), so it reads as the same latent grid at every size.

• Canvas 2D only, no WebGL and no per-pixel work per frame: a few drawImage calls. Noise is generated once per step at latent resolution (a few thousand cells).
• The prediction chain is built once per image and size; decode tiles are direct copies from the sharp image.
• Frames stop drawing when the image is resolved and nothing changes, and the loop stops entirely off screen.
• Images are decoded off the main thread before use.

• Use for image generation surfaces: an image canvas, an inline image in a chat reply, a variations grid, or an empty state while a first image renders.
• Wire a real job by passing status and progress (step / total), plus the job's seed. Pass decoding while the VAE decodes if the backend reports it. With no progress events, pass status only and an estimated duration.
• Use size="sm" in grids. A batch samples in lockstep, so give every tile the same duration and its own seed, and override max-inline-size through className or style.
• The preview is sampled from the final image, so pass src as soon as a preview or final URL exists; until it loads the tile stays in the calm queued look.
• For text or agent work in progress, pick a text shimmer or an agent status component instead.

The full library index for assistants is at /llms.txt.