Case study · Images, video and DAM
Every image on the right variant
In short
Product images for the sports catalogs sat in enterprise DAMs and media hubs, with no clean link to the product, the color or the angle. alfa built a harvester that reads each type of server, removes duplicates and ties every image to its color and angle, then puts the hero image of each color on every variant row. Galleries match the variant the shopper picks, with no manual sorting.
The problem
Brands rarely hand over a clean folder of images. The images sit inside closed DAM systems, each with its own addressing, its own image transforms and its own angle names. Downloading them by hand at catalog scale produces mismatched colorways, missing sole and profile views, and broken variant links.
What alfa built
One harvester, with a reader for each type of server.
Scene7 does not serve fixed files. It renders each image on request from the URL. The harvester builds the URL for each asset from the SKU, probes every angle code the server uses, and sets width, quality and sharpening in the same request for a clean export.
The assets were filed under internal IDs, not product codes, and the hub had no bulk export. The harvester walked the collections, matched each internal article ID to the brand's ERP product code, removed identical hero images by perceptual hash, and grouped a full gallery for each colorway.
Desktop, mobile and lifestyle images sat in separate trees behind Akamai and Cloudflare. The crawler mapped each colorway swatch to its full zoom set, checked the size of every response, and placed the images on the Shopify CDN tied to their variants.
A second look: vision check (beta)
Harvesting answers where the image is. The next question is whether the image matches the color name the supplier gave it. A vision model turns each cropped product image into a vector, and the vector is checked against the record.
- Colorway check: the dominant colors are compared with the variant's color name, so a green variant mapped to a red shoe is caught before it ships.
- Metal classification for jewelry and luxury goods: yellow gold, rose gold, stainless steel, platinum and matte titanium, read from the highlights in the photo.
- Attribute tags: material and finish written as filter values from the image, not typed in by hand.
Rules it left behind
- The hero image of each color goes on every variant row.
- Video goes in as a gallery row.
- Never de-duplicate by file-name family.
- Image sources are always full URLs, and every one is probed before import by the validation gate.
The result
Galleries that match the variant the shopper chose, with no manual sorting. The harvester ran on the catalogs behind Under Armour Israel and Saucony Israel, both live stores.