We replaced a week-long human editing workflow with an automated AI pipeline turning three raw shots of an eyewear frame into eight studio-grade, ERP-ready variants in under an hour.
The client's e-commerce pipeline ran on a manual editing queue raw studio photography in, eight format-perfect assets out, with humans doing every step in between.
Three raw shots per frame (front, diagonal, side) were handed off to human editors who corrected angles, balanced lighting and shadows, removed backgrounds, and standardized aspect ratios producing a white-background JPG and a transparent PNG for every variant. Lead time for a single frame regularly stretched to a full week.
Build an automated, AI-driven pipeline that matches the visual quality of expert manual editing correct lighting, clean mattes, accurate half-glass crops at a fraction of the cost and turnaround time, while elevating the consistency of the catalog overall.
Each capability was designed to be operable by the catalog team not just the ML org. Together they take a raw shot to an ERP-ready asset without a manual touch.
Gemini-driven enhancement normalizes lighting, contrast, white balance and aspect ratio across batches matching the look of the in-house studio standard.
An advanced rembg model (isnet-general-use) extracts pixel-accurate transparent masks without nicking thin temples, hinges, or reflective lens edges.
Every upload fans out into JPG, PNG, and WebP across transparent, white-background, and half-glass variants up to eight ERP-ready files per source image.
Custom vertical-projection bridging logic detects the eyewear's bridge and crops a precise half-glass variant for angled shots automating a previously eyeballed manual task.
Celery workers backed by Redis handle AI inference, compositing, and S3 transfers in parallel, with the Gemini Batch API used for non-urgent jobs at 50% cost.
JWT-authenticated frontend with strict RBAC ensures only authorized catalog operators can upload, queue, or download generated assets.
We picked components the client's engineering team could own after handoff proven web stack, contained model surface, observable workers.
Three places the textbook approach broke and what we shipped instead.
The bar wasn't just "automated" it was "indistinguishable from an expert editor." That meant nuanced lighting correction, clean lens reflections, and precise edge detection on delicate temples, all without a human in the loop and at a fraction of the cost and turnaround.
We assembled a multi-stage processing pipeline: Gemini AI handles lighting, shadow, and aspect-ratio normalization first, then an advanced rembg model (isnet-general-use) extracts a transparent mask without damaging thin frame elements. The two stages compound neither alone matched the standard; together they did.
The catalog required a "half-glass" crop variant of angled shots a view editors previously produced by eyeballing the bridge of the frame and cropping freehand. There was no off-the-shelf model for it, and frame geometry varies wildly across SKUs.
We wrote custom bridging logic using vertical projection over the alpha mask finding the bridge column algorithmically and cropping a precise half-glass view. A subjective manual task became a deterministic, repeatable function.
Catalog drops meant hundreds of high-resolution RAW images arriving at once. Synchronous handling would have blocked the API, starved the workers, and stalled S3 uploads and the client couldn't accept a queue that fell over under peak load.
We split the HTTP layer from background processing. FastAPI accepts and records jobs instantly; Celery workers on Redis handle AI inference, Pillow compositing, and S3 transfers in parallel. Non-urgent jobs route through the Gemini Batch API at 50% cost same quality, longer SLO.
Five stages from RAW upload to ERP-ready asset.
Catalog operators authenticate via JWT with strict RBAC, then upload batches of RAW images through a React + Vite frontend. The backend parses the strict filename convention ({PRODUCT_CODE}_{VIEW_NUMBER}.ext) and auto-detects view type 1=Front, 2=Angle, 3=Side.
FastAPI writes job metadata to PostgreSQL (asyncpg) and dispatches a Celery task. The HTTP layer responds with an accepted state immediately so the frontend can poll without UI blocking. Workers run a separate sync DB driver (psycopg2) to keep responsibilities clean.
The worker streams the image through the Google Gemini API to correct aspect ratio and perform studio-grade enhancements lighting, contrast, white balance. Non-urgent jobs route through the Gemini Batch API at 50% cost for the same quality.
rembg (isnet-general-use) produces a pixel-perfect RGBA mask. Pillow composites the standard white-background variants. For Angle views, a custom vertical-projection routine finds the bridge and crops a half-glass variant programmatically.
The pipeline generates an in-memory matrix JPG / PNG / WebP × transparent / white-BG / half-glass and streams up to eight assets per source image to S3. PostgreSQL records the artifact set; an S3 lifecycle policy expires assets at 60 days and a Celery Beat cron cleans the DB rows.
Numbers measured against the pre-rollout manual workflow. The system replaced a tedious editor queue with a seamless AI engine that matches human quality at a fraction of the cost.
What used to take a full week of manual editor handoffs now completes in minutes; automated end-to-end, with no editor in the loop.
JPG / PNG / WebP × transparent / white-BG / half-glass generated dynamically per source image.
AI processing costs a fraction of manual output with further savings on non-urgent jobs via the Gemini Batch API.
Catalog QA accepts AI output at parity with the previous human-editor baseline.
Lifecycle policy expires assets; a Celery Beat job cleans the DB rows in lockstep.
The pipeline matches what our editors used to ship and it does it before the next batch even lands.
310 Kuber Avenue, Near Gurudwara Cross Road, Jamnagar – 361008
Plot No 36, Galaxy Park – II, Morkanda Road,
Jamnagar – 361001
Copyright @Azguards Technolabs 2026 all Rights Reserved.
