Slide dither loader

One renderer for the gap between "we asked for an image" and "the image arrived", on any slide. It takes three things it can already know at render time, the slide background, the accent colour, and the shape the image will occupy, and screens that shape in dots until the picture lands.

The three slides below share a single function. Nothing about them is special-cased: a full-bleed rectangle on white, a chevron on dark navy, and a folded polygon on cream all come from the same call with different arguments.

Three slides, one call

The contract

Why it is built this way

• The region is tinted, not outlined. A stroke and an inner glow drew the container's edge; a low-opacity blue field with a few very large gradients drifting through it says the same thing without a line. The mask is felt at the tint boundary rather than traced, which suits an angular shape better, and the slow drift reads as "working" without a spinner. Stroke and glow are still there at 0 if you want to compare.
• Box and shape are separate, because that is how the product models it. In app.presentations.ai the picture is a .ci-element-image placed in the slide, carrying a clip-path written in percentages of its own box. The renderer takes the same two values, so a mask can be lifted from a layout and dropped in here unchanged.
• The shape is the source of truth, used twice. The same percentages become both the clip-path on the arriving picture and a Path2D for the dot test, so the loader and the image can never disagree about where the image goes.
• These are the live values, not a trace. Every box, clip-path, background and picture below was read out of the three decks in app.presentations.ai, so the geometry is exact rather than eyeballed. Deck ids are printed on each slide. The Daniels chevrons are the real thing too: the mood paints a solid accent masked by its own PNGs (line-tr1.png, line-br1.png), which is why its gold is #FFBA08 and not the softer tone a screenshot suggests. The Instagram blocks and the Lenskart purple folds are rebuilt by eye — that deck's mood has since been changed in the app, so its original decor is no longer there to read.
• Dots are placed by centre, not clipped. Every cell centre is tested against the shape and skipped if it falls outside. Clipping the canvas instead would slice dots in half along the edge, which reads as a mistake on an angular mask.
• Ink colour is derived, then checked. The accent is used as the dot colour, but only if it clears a contrast ratio against the slide background. A navy accent on a navy slide falls back to a tint of the background, so the loader can never come out invisible.
• Canvas, not DOM. A full-bleed region at pitch 8 is around 7,000 cells. As elements that is a stutter; as one canvas with dots batched into eight alpha buckets it is a few fills per frame.
• White noise stops working at this size. Over a few hundred cells a random per-cell phase reads as lively. Over seven thousand it averages out and the whole region looks uniform, because at any instant every phase is equally represented everywhere. The fix is spatial coherence: neighbouring cells need to share a phase so the eye can see a patch. That is what Clumped, Drift, Rings and Bands are — a smooth field instead of noise, with Patch size setting how big a patch is and Coherence blending back toward pure noise. Random (white) is kept so the difference is visible.
• Three phases, because the first one cannot know the picture. At t=0 the image does not exist yet, so a screen that already outlines the subject is lying. Phase 1 is a neutral field. Phase 2 begins when a preview lands: preview(toneFn) morphs every dot from the field toward the picture's own tone, so the noise reorganises itself into a low-resolution render. Phase 3 swells the dots out as the full picture fades in. Only the middle phase knows anything, and it only knows what a preview could actually tell it.
• Tone is pluggable, and here it is plugged in. Switch Tone to Neutral field and the screen is decorative, which is all you can do when the picture is genuinely unknown. Leave it on From image and imageTone() samples the real photograph, so the dots are a low-resolution render of the picture they resolve into. In production that sampler would read a tiny preview from the generator rather than the full image.
• The sampler matches the crop, and follows the ink. It applies the same cover crop the CSS will, so the preview cannot be framed differently from the photo. And tone means darkness on a light slide but brightness on a dark one, because the ink flips. Without that, the halftone renders as a negative.