Built a Figma plugin for removing objects from images with AI. First version took 12 seconds. Nobody waited. Got it down to 2 seconds.

What was slow: - Base64 conversions (Figma's image API is clunky) - Wrong AI model (Stable Diffusion was overkill) - Waiting for complete response before showing anything

What worked: - Switched to LaMa model (faster, simpler) - Show preview immediately, refine in background - Better use of Replicate's streaming API

Figma plugin quirks: - Can't edit pixels directly - export/import only - Memory limits crash on large images - Transparent PNGs break everything

Got 105 users in a week from 2 Reddit posts. Currently averaging 1.8 seconds per removal.

Figma Plugin: https://www.figma.com/community/plugin/1576512610054427811/p...

Replicate Model: https://replicate.com/dpakkk/image-object-removal

Anyone else optimized plugins in sandboxed environments? What tricks am I missing?