Volume against strict turnaround
High image volume with tight SLAs required structured batch workflows and defined daily throughput targets.
Client success story · CS28
25,000+ AI-generated images, audited against a written rubric
Visual consistency validation for a large multi-category retailer
- IndustryE-commerce / home furnishings & lifestyle
- Catalog scale100,000+ active SKUs
- ScopeSofa products, full visual consistency audit
- PlatformAWS SageMaker
- 99%annotation accuracy rate sustained
- 38%reduction in image rejection rate
- 25,000+images processed
Client background
A large-scale multi-category retailer across home furnishings, lifestyle and interior décor, with an active catalog exceeding 100,000 SKUs, had invested heavily in generative AI tooling to produce product variants at scale.
AI generation introduced a new category of risk. Without a validation layer, generated images were reaching the catalog with warped geometry, colour drift from the reference image, backgrounds violating brand guidelines and structural artifacts.
Problem
Operational challenges
For a business where the product image is the primary trust signal, and where a customer cannot sit on a sofa before buying it, visual consistency is commercial rather than cosmetic. A single inconsistency in colour, geometry or background erodes confidence, inflates returns and degrades catalog integrity.
The programme had to validate every generated image against a clearly defined standard, at the volume and speed the catalog operation demanded.
Consistency across a large team
Maintaining annotation consistency across 25+ people needed a standardised SOP with visual examples per defect category, daily calibration and inter-annotator agreement tracking.
Edge cases
Modular sofas, detachable chaises and irregular orientations needed dedicated guidelines with illustrated decision trees — L- and U-shaped sofas received flexible pose rules, detachable components assessed for structural plausibility rather than exact replication.
Borderline subjectivity
Explicit rule-of-thumb logic was embedded in the SOP: any element not explicitly allowed was treated as forbidden, with borderline cases escalated to a senior QA reviewer within the same shift.
Solution
Custom solution executed
-
01
SOP-driven workflow
Detailed SOPs covered all quality parameters with clear criteria and visual examples, so image evaluation was consistent and objective rather than dependent on individual judgement.
-
02
Three-image comparison
Each annotator worked with a three-image set — the main product image as reference, the image variant as secondary reference, and the generated image under evaluation — using both references to assess geometry and colour, texture and material compliance so no single reference point introduced bias.
-
03
Rule-based validation
Each generated image was assessed against ten structured quality dimensions with explicit pass/fail criteria: geometry, visibility, colour accuracy, material fidelity, pose, background compliance, object count, allowed versus forbidden elements and artifact detection.
-
04
Error tagging and defect taxonomy
Non-compliant images were labelled with standardised defect categories — geometry artifacts, wrong orientation, background issues, texture and colour issues, generated-to-main mismatch — enabling tracking and reprocessing, and giving the client's AI team attributable failure modes rather than a rejection count.
-
05
Multi-level QA
Each image was independently audited by three different auditors, with further sampling validation by senior QA specialists and full re-review triggered when error rates exceeded predefined thresholds.
Workflow
Execution pipeline
Every image moved through six controlled stages inside AWS SageMaker, giving end-to-end traceability from batch intake to production-ready delivery.
- 1Image ingestion Batches of AI-generated images were received and managed within SageMaker, then systematically mapped to their corresponding main product images and variants for structured evaluation.
- 2Reference comparison Annotators reviewed the three-image set — main, variant, generated — inside the SageMaker interface to establish a baseline for geometry, colour and material consistency.
- 3Rule-based validation Each generated image was assessed against the predefined SOP criteria: geometry, visibility, colour accuracy, material fidelity, pose, background compliance, object count, allowed versus forbidden elements and artifact detection.
- 4Error tagging Non-compliant images were labelled directly within SageMaker using the standardised defect categories, so every rejection carried an attributable reason and could be tracked and reprocessed.
- 5QA review Multi-level quality checks: each image audited by multiple reviewers, then further validated through sampling by senior QA specialists.
- 6Final approval and delivery Fully validated images were approved and delivered as production-ready outputs, alongside structured defect reports and taxonomy insights supporting the client's re-generation and continuous improvement workflows.
Setup
Platform and team
- PlatformAWS SageMaker
- TeamTrained annotators supported by QA leads
- Quality modelThree independent audits per image, plus senior sampling
- ScalingStandardised tools and workflows, so volume grew without loosening the standard
Impact
Business value delivered
- 99% annotation accuracy sustained across a 25,000+ image dataset
- 38% reduction in rejection rate against the pre-engagement baseline
- 25+ specialist annotators deployed, each sustaining accuracy above 98%
- Top three generation failure modes identified and resolved by the client's AI team within the first month
Impact
What it changed for the shopper
The audit was commercial rather than cosmetic, and the document records its effect on the buying decision itself.
- Higher trust in product visuals translates directly to lower pre-purchase abandonment
- Consistent, professional imagery across the sofa catalog reinforces brand quality perception
- Accurate colour and material representation reduces post-purchase disappointment and return initiation
What it changed for the client
- Product publishing accelerated by removing the manual review bottleneck at catalog intake
- Catalog visual consistency measured against a documented, auditable standard
- Reduced rework volume through clearer defect attribution
In short
Generative tooling changed the cost of producing product imagery and not the cost of getting it wrong. The value here was the validation layer: a written rubric, three independent audits per image, and defect tagging structured so the client could fix the generation model rather than keep rejecting its output.
Services this story used
Related case studies
Have a version of this problem?
Describe it in a sentence. We will tell you what we would do about it, and whether it is something we should be doing at all.