A product catalog is the store. When an entry is a single low-resolution image and one line of text, search cannot find it, filters cannot narrow it, and a shopper cannot judge it. That gap is where engagement and conversion are lost — before anyone reaches checkout.
We built a pipeline that reads a bare product image and returns a full, locale-aware catalog entry: categorized attributes, translated titles and descriptions, culturally-appropriate imagery, and an interactive 3D asset. Every generated output passes an automated quality gate before it ships. This paper describes the architecture, not any one model behind it.
1. The problem
Catalogs decay in predictable ways.
- They start thin. A supplier hands over a photo and a SKU. The rich attributes that power search — material, pattern, silhouette, use case — never get entered.
- Human categorization is inconsistent. Two people tag the same item differently. The same product lands in different categories across regions, and faceted search fails quietly: the item exists but cannot be found.
- They go stale. Terminology shifts, seasons turn, and last year's descriptions stop matching how customers actually search today.
- Managers are blind to real usage. There is no feedback loop from how a product is described to how it is found, so nobody sees which terminology resonates or what is trending.
Manual enrichment does not scale against this. It is slow, expensive, and its inconsistency is the very thing that degrades search. The work has to be automated, and the automation has to be trusted enough to run without a human reviewing every field.
2. Pipeline overview
The pipeline is a sequence of specialized stages, each owning one role. An input image enters; a structured, multi-locale, multi-format entry leaves. A reflection stage sits across the generative steps and can send any output back for another attempt.
3. Product analysis and categorization
The first stage is a vision-language model that looks at the product image and describes what it sees. It extracts the attributes a shopper would filter on — material, colour, pattern, form, apparent use — and grounds each one in the pixels rather than guessing from a sparse title. This is where a thin entry gains the structured data that search and filtering depend on.
The same stage performs categorization: the extracted attributes are mapped onto our predefined category taxonomy. Because one model applies the taxonomy consistently, the drift that comes from many hands disappears. The output is a normalized attribute record — the spine every later stage reads from.
4. Culturally-aware prompt planning
Generation is only as good as the instruction it is given. Between analysis and generation sits a reasoning model that plans the prompts. It takes the attribute record and the target locale and decides what a believable, appropriate presentation looks like there: which settings, palettes, props and framing read as native rather than transplanted.
This is the stage that prevents a generic global background from being pasted behind every product. A homeware item destined for one market might be staged in a regional courtyard; for another, a formal interior; for another, a shared family space. The planner encodes that intent as concrete prompts, and it does so per locale, so the same product is presented in terms each market recognizes.
5. Multi-locale titles and descriptions
The attribute record drives copy generation across roughly ten locales. We do not translate one canonical description; we generate for each locale from the structured attributes, so the result reads as native copy and uses the terminology of that market rather than a literal rendering of another. Titles stay within catalog constraints; descriptions surface the attributes that matter for discovery. Brand voice is carried as a constraint into every locale so the catalog stays consistent across languages.
6. Localized image generation
An image-generation model produces product variations using the plans from stage 4. The product itself is held fixed — recontextualization changes the setting, not the item — while backgrounds and staging are rendered to suit each locale. The output is a set of on-brand, culturally-appropriate images per market, generated from one source photo instead of commissioned per region.
7. From 2D to interactive 3D
For products that benefit from inspection, an image-to-3D model lifts the source photo into an interactive 3D asset. We export the standard binary glTF container (GLB): geometry, textures and physically-based materials packaged for real-time web viewers and AR. A shopper can rotate the item and judge form and finish — the closest thing to handling it — from a single input image.
8. Automated quality assessment
Generative output cannot be trusted blindly, so quality control is part of the pipeline rather than a manual step after it. A vision-language reflection stage scores each generated image against the intended attributes and the planned presentation — is this the right product, is the setting appropriate, is the result clean and on-brand.
Scoring closes a loop. When an output falls below threshold, the reflection stage identifies the weakest dimensions and feeds targeted corrections back to the planning and generation stages, which retry. Only outputs that clear the bar are published. This is what lets the pipeline run at catalog scale without a reviewer on every asset: the judgment is automated and applied uniformly.
9. A modular API surface
The stages are exposed as separate endpoints — analysis, image generation, 3D generation — rather than one opaque call. A team can enrich attributes on an existing catalog without generating imagery, add localized images to products that already have copy, or attach 3D only to a hero range. Each stage scales independently, and the reflection loop can be tightened or relaxed per stage.
10. Key capabilities
- Grounded product understanding — attributes read from the image, not inferred from a sparse title.
- Consistent categorization — one taxonomy applied by one model, ending human drift.
- Locale-aware planning — presentation designed per market before anything is generated.
- Native multi-locale copy — titles and descriptions generated, not translated, across ~10 locales.
- Culturally-appropriate imagery — product held fixed, setting localized.
- Interactive 3D — a single photo becomes a rotatable GLB asset.
- Automated quality gate — a reflection loop that scores, corrects and retries before publishing.
- Modular endpoints — analysis, imagery and 3D consumed independently.
11. Deployment requirements
Generalized, since the specific models are interchangeable:
| Requirement | Detail |
|---|---|
| Compute | Multiple GPUs to self-host the vision-language, generation and 3D models concurrently |
| Storage | Scalable object storage for variations, 3D assets, metadata and quality artifacts |
| Runtime | Containerized services, one per stage, scaled independently behind the API |
| Throughput | Sized to catalog size and locale count; reflection retries add headroom per item |
12. Ethical considerations
Trustworthy AI is a shared responsibility. Generated imagery and copy represent real products in real markets, so cultural presentation is reviewed for appropriateness rather than assumed correct, and the quality loop guards against misrepresentation. Teams should confirm each stage meets the requirements of their industry and use case, and address foreseeable misuse before publishing at scale.