Shotsmith

Small e-commerce sellers need many images per product: a white-background main image for Amazon, lifestyle shots for Instagram, square crops for Shopify and banners for their site, plus listing copy. Studio photography and manual editing for every product is slow and expensive.

Shotsmith had to turn one raw product photo into all of that, without distorting the product or its label, while keeping AI costs under control and never charging users for a generation that failed.

The studio is a Nuxt 4 app with a responsive canvas: zoom, pan, a transparency checkerboard and a before/after slider. The Magic Edit brush is built on vue-konva, with brush sizes, an eraser, an undo stack and a natural-language instruction for what to change in the masked area. A results strip shows the newest outputs first with lineage, progress and elapsed time.

Background removal uses fal.ai's BiRefNet through its queue, with webhooks and polling, and falls back to a local Sharp engine that flood-fills from the edges for zero-cost development runs. Scene generation calls Gemini image models to produce four variations per run from presets or a custom prompt. Gemini Vision describes the product and writes a title, description, five Amazon bullets, an Instagram caption and hashtags in several tones and languages. ESRGAN handles 2× and 4× upscaling.

Every generation runs as a durable Inngest job with retries and step recovery. Credits live in an atomic ledger in Neon PostgreSQL with Drizzle: they are deducted before a job and refunded automatically if it fails. Exports are produced with Sharp for Amazon Main (2000×2000, white, 85% fill), Instagram post and story, Shopify square and web banners, and zipped with archiver. Upstash Redis rate limits, Better Auth sessions and a strict Content Security Policy protect the app.