Image & content pipeline

A B2B furniture catalog lives on photos, and most of them are taken in clients' offices: wrong light, brick columns, other people's furniture in the frame. Reshooting every product in a studio is not an option for a small manufacturer. The pipeline takes the real photo and rebuilds the environment around the same product.

Generation takes minutes, and a batch job up to a day, so the queue message carries only the task id — the worker reads the state from the database and continues from where it stopped, which makes redelivery safe. Temporary Gemini errors go to a wait queue with a TTL and come back (up to three attempts); poison messages go to a dead-letter queue. One or two frames are generated by ordinary calls, three or more through the Batch API, which is twice as cheap. Every model call is written to a cost log with an estimate in dollars.

On first save the EXIF rotation is applied to the pixels and the metadata is stripped. The page, not the upload API, decides the collection and the file name — a product photo is named by type, model and index. Renditions are named by their actual size and served as a WebP srcset. For ads a 4:3 JPEG is cut from the original so that the crop is slightly larger than the target and gets scaled down to exactly 4:3 — small originals would otherwise round to "almost 4:3", which the ad platform rejects.