AI Agent · E-commerce
Favia Decor
An AI batch agent that turns raw product photos into store- and social-ready imagery.
Overview
Commerce Image Agent is an AI-powered batch tool that turns raw, inconsistent product photos into polished, e-commerce-ready and social-media-ready imagery automatically. It analyzes each product photo, understands what the product is, and regenerates it as a clean studio or lifestyle hero shot, then formats branded variants for online stores and Instagram. It was built and deployed for Favia Decor, a home-decor brand.
The challenge
Small and mid-size e-commerce sellers shoot hundreds of products on phones and point-and-shoot cameras, producing uneven lighting, cluttered backgrounds, and off-brand framing. Hiring photographers and retouchers per SKU is slow and expensive. The brand needed to automate category detection, enhancement, brand overlay, and channel formatting across an entire catalog in a single run.
What we built
- A multi-stage pipeline where a Gemini Flash vision model analyzes each image to detect product category and write a per-image creative brief, then a native image-generation model regenerates a clean hero, mirroring the original folder tree into an e-commerce-ready output.
- A model "waterfall" with quota resilience: analysis falls back across Gemini 2.5/3 Flash and Flash-Lite, generation uses Gemini 2.5 Flash Image with an Imagen 4 fallback, and smart HTTP 429 handling parses the API's exact retry delay with exponential backoff, configurable concurrency, and skip-existing resume safety.
- An Instagram launch module that generates scroll-stopping social creatives at an exact 4:5 ratio and composites the client's real brand logo on top, done deliberately, since AI cannot reproduce an exact brand mark reliably.
- A multi-model cost and quality comparison harness that runs the same product through several Gemini, Imagen, and OpenAI gpt-image variants, then produces a JSON, CSV, and HTML gallery report plus a PDF, with per-image pricing pulled from a configurable, currency-aware pricing file.
Results
The agent collapses per-SKU photo retouching and restaging into an automated batch run across a whole catalog, producing consistent, on-brand studio and lifestyle imagery from amateur source photos. A single source photo yields multiple channel-ready assets, a store hero plus a branded Instagram post. The built-in model comparison lets the team choose the best quality-per-cost model before committing to a full run, and resumable processing avoids re-spending on already-processed images.
