Turn pictures into sprites with a no-AI subject finder

[Finesssee]

Viewers who make their own sprite on the fighter page ("Your own sprite") get a cleaner cut-out without any AI call.

What changes

• New src/autosprite.ts finds the subject: border background colors, geodesic saliency, a body prior for busts, and a graph cut. A picture with no clear subject (a landscape, two animals side by side) keeps the whole picture instead of a random piece. A blank picture still gives nothing.
• src/spritify.ts keeps its API. Shrinking now uses a k-means Oklab palette (14 colors), weighted cell votes so thin dark lines and eyes survive, a lone-pixel cleanup, and an outline in a dark shade of each border color.
• Removed palette, quantize, outline and OUTLINE (no callers outside the tests).

Checked

• bun run check: 391 of 391 pass. tests/spritify.test.ts covers a scene keeping the whole picture, the shaded outline literals and the existing cut-out and size cases.
• Bench on seven real pictures (cartoon bust, emote sheet, three duck busts, landscape, lion and cat): 150 to 560 ms each in Node.

Limits

• It runs on the main thread, so the page pauses for up to about 0.5 s while a big picture converts. A worker could follow.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Configuration used: Repository: nesszer/pixfray/.coderabbit.yaml
• Review profile: CHILL
• Plan: Advanced
• Run ID: 442954ed-b51c-4628-86df-f123166efd7a

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}