A quick-commerce catalogue team made product-image processing repeatable

01

Business context and previous workflow

Catalogue teams receive product data and images from many sources. Images can have inconsistent backgrounds, sizes, aspect ratios, and file quality. A manual process requires a user to download each image, prepare variants, select a result, resize it, upload it, and update the catalogue file.

The wider programme also considered missing product attributes, content quality checks, visual enrichment, and later media formats. Those items appear in a statement of work. They are not all established as delivered.

02

Implemented image workflow

A user prepares a CSV with a product identifier and one or more image links. A Cloud Run job downloads the images. It creates two background-removed variants and uses Gemini to select the better result. It then applies the required aspect ratio, writes the files to Cloud Storage, and produces an output CSV with links to the original, intermediate, selected, and final images.

The user can process the first image for each product or all available image columns. Large files can use Cloud Run tasks to divide the input across shards. One task waits for the parts and merges the result when all required outputs are present.

• CSV input with product and image references
• Two background-removal variants
• Gemini-assisted selection between prepared variants
• Aspect-ratio processing and Cloud Storage output
• Output CSV and job monitoring

03

Operating checks and human review

Operators check that source URLs are available, tasks complete, output files exist, and the final image matches the product. A failed shard must prevent a silent incomplete merge. Logs and execution status support diagnosis.

Human review is still required for product accuracy, text on packaging, colour, damage repair, and brand presentation. Model selection does not prove that an image is fit for publication. The catalogue owner decides the acceptance rules and release process.

04

What remains proposed

The statement of work proposes attribute generation, content checks, confidence flags, image repair, lifestyle edits, video, translation, size-chart conversion, and other later functions. It also lists quality and throughput targets. This story does not present those roadmap items or targets as completed results.

The implemented guide establishes a repeatable image workflow and an operating method. It does not contain an approved measure for time saved, quality improvement, catalogue conversion, or production volume.