Drop Studio

CASE STUDY

Model photos for a clothing drop used to mean one manual image-generation chat per item. Drop Studio takes a style reference and a folder of outfit photos, and produces a reviewed, watermark-free set of listing images in one run.

View the code on GitHub →

20–30 passes → 1 clickone manual chat loop per item, replaced by a single run

3 at a timebounded queue, per-photo retries

~4.5¢ / imageon the default engine, per published pricing

01 · The problem

The same manual loop, 20 to 30 times per drop

Producing model photos meant repeating one loop for every item: open an image-generation chat, upload a style reference, upload the outfit, paste a photographer prompt, wait, save, repeat. A drop is roughly 20 to 30 items. The consumer chat apps also stamped a visible watermark on every output.

02 · How it works

Reference in, listing-ready zip out

Stage 0 · Client

Intake

One style reference plus any number of outfit photos. iPhone HEIC converts to JPEG in the browser; everything is resized and compressed before upload.

Stage 1 · Modes

Pick a mode

New Setting takes scene and lighting from the style reference. My Models reuses a saved model, so one face stays consistent across a store.

Stage 2 · Queue

Generate

One click runs the whole drop through a queue that generates three at a time, streaming each finished photo into a results grid.

Stage 3 · Export

Review & export

Regenerate any single photo, keep what works, export the drop as a zip named by outfit, ready for product listings.

One failed generation shows an inline retry and never kills the batch. Provider keys live only in server environment variables, and a passcode gate keeps strangers from spending API credits on a public deployment.

03 · Design decisions

Reviewed by a human, engine-agnostic by design

Human in the loopEvery image reviewedI review every photo, regenerate individually, and choose what to keep. Nothing is auto-published.

ArchitectureProvider-agnostic engine layerThe server normalizes each model's request and response behind one interface, so image engines swap from a dropdown without touching the UI.

04 · Results

Validated end to end, honestly reported

A real generation has run through the full pipeline; the example output lives in the repo. A full production drop has not been timed yet, so time saved versus the manual process is not measured. Cost on the default engine is about 4.5 cents per image based on published pricing, and API-returned images carry no watermark, unlike the consumer chat apps this replaces.

05 · Under the hood

Stack

Next.js · React Node serverless image proxy Gemini + GPT image models via API heic2any · canvas · jszip Vercel · PWA

Public code, keys and store data excluded, on GitHub. Built for Transcend Vintage.