Visual search built on embeddings of your own catalogue
Visual search works because images can be turned into numbers that put similar-looking things close together. We encode every catalogue image with a CLIP-style model into a vector, store the vectors in an index such as Qdrant or pgvector, and encode the shopper's photo the same way at query time, so finding the closest products is a nearest-neighbour lookup that returns in milliseconds across a large catalogue. A query photo usually contains a room, a hand or a shop shelf as well as the product, so a detector crops to the item first. A generic embedding ranks by whatever it finds salient, often colour, when your shoppers care about cut or material, so we fine-tune on your catalogue until the distances follow your taxonomy. The same embeddings drive similar-item recommendations and the automated tagging behind your filters.
Fine-tuning image generation so products stay accurate and on brand
The risk with generated product imagery is not that it looks bad but that it looks plausible and wrong, with a buckle moved or a colour shifted on an item a customer will hold. We handle this by separating what the model may invent from what it must preserve. A LoRA fine-tune on your own photography teaches the base diffusion model your lighting, palette and composition, while ControlNet conditions each render on the real product's edges and depth, so the shoe in the lifestyle scene is your shoe. For banners and social creative the constraints loosen and each render costs seconds of GPU time. Two conditions apply. You must hold the rights to the images we train on, and every published render keeps its prompt, seed and model version on file, so a creative can be reproduced or withdrawn later.
Running visual search and image generation in production storefronts
In production the index must track the catalogue without anyone remembering to retrain. New SKUs are embedded as they arrive through your product feed or a webhook, discontinued lines are dropped from the index, and stock and price filters are applied inside the vector query so a shopper is not shown a perfect match that cannot be bought. Search endpoints are served from ONNX Runtime or TensorRT behind FastAPI with a latency budget agreed up front, while generation runs on a separate GPU pool that scales to zero overnight, so the bill tracks use. We measure what the storefront cares about, holding back a share of traffic without visual search so the lift in click-through and add-to-cart is real rather than seasonal. You keep the fine-tuned weights, the index and the dashboards, documented so they can move with you.