Building a Constraint-Based Image Pipeline for Amazon Listings

#ai #design

I recently replaced eight separate image briefs with a single constraint system for Amazon listing image generation. Here's the technical thinking behind it, the pipeline architecture, and the QA gates that catch AI drift before it reaches a buyer.

The problem: AI image generators have no memory of the real product

When you generate an Amazon listing image set in six separate sessions with a standard AI tool, each session is amnesiac. The main image has one lid shape. The detail shot has a slightly different one. The lifestyle scene invents a gloss finish the real product doesn't have.

This isn't a prompting problem. It's an architecture problem. The tool doesn't maintain a constraint object across generations.

What I built instead using LoomaDesign's pipeline

The system takes one structured product truth object and fans it out across image types. Think of it as a constrained generation pipeline:

The pipeline order is deterministic

Stage order matters more than any single prompt:

Source preparation — white background removal while preserving edge geometry
Enhancement gate — sharpness recovery only if original detail is soft; reject if it changes product shape or color
Gallery expansion — angle, detail, scale, lifestyle from the locked reference
A+ module construction — reuse the same product truth object, different layout templates
QA pass — compare every output against the constraint object
Each stage inherits constraints from the previous one. This is the key architectural decision: the pipeline is append-only for constraints, never destructive.

The QA program is more important than the generation

Most of the real work is rejection logic:

If an image fails any gate, it's rejected. The pipeline moves to the next attempt with the same constraints, not a looser prompt.

Scale and performance

For a 7-image set (main + 6 support images + A+ modules), the pipeline runs in connected passes rather than independent sessions. Traditional workflow: 6-8 separate briefs, each requiring context rebuild. Under constraint-based generation: one structured input, fan-out with inherited constraints, sequential QA.

The takeaway isn't "better prompts." It's that reliable AI image generation for ecommerce requires a constraint system, not a prompt library. The prompt changes every time. The product constraints shouldn't.