Automating Product Imagery, Saving Millions In Production Costs.
How I helped a luxury sculpture studio replace much of its photography, interior-design and physical staging workflow with an AI production system — so 1,000+ products could be marketed without manufacturing and photographing every piece first.
Why the catalogue could not be photographed the old way. Seven physical stages stand between one product and one usable image — and the whole chain repeats for every product, every view, every time.
1,000+
High-value catalogue products
~$30–40K
Typical product value
~7
Coordinated marketing views per product
The Engagement
Building an AI photography system for a 1,000+ product catalogue.
A Dubai luxury sculpture studio had more than a thousand high-value products to market, many selling for $30K–$40K. Buyers needed to see those pieces inside real architectural environments — villas, hotels, courtyards, spas, commercial spaces — not isolated against white backgrounds.
Traditionally, producing that imagery meant manufacturing or preparing the physical product, transporting and installing it, securing premium locations, styling the environment, hiring a photography crew, shooting multiple views and retouching the assets. At catalogue scale, product marketing became a multi-million-dollar physical-production problem.
I was brought in to determine whether generative AI could replace enough of that workflow to make catalogue production commercially scalable. What started as a ten-product experiment became a production system combining LLM-based planning, multimodal evaluation, catalogue retrieval, client-preference memory, product and scene guardrails, confidence-based routing, human review, automated QA and publishing.
The result was not an image generator. It became an AI photography and interior-design production workflow.
Client Constraint
The cost wasn’t taking the photograph. It was everything required to make the photograph possible.
A sculpture can only be understood properly when its scale, material and placement are visible in context. For every product the client needed several views: scale in a real room, architectural placement, material under natural and artificial light, alternate angles, close detail, and day against evening.
At roughly seven useful views per product, a catalogue of 1,000+ products implied more than 7,000 finished marketing images. Producing those conventionally meant repeating an expensive physical process for each one — fabrication, transport, installation, location, styling, photography, retouching.
The client needed high-quality environmental imagery to sell the products, and could not justify physically staging every product purely to create it.
Could I recreate enough of the photography and interior-design workflow in software without compromising the physical truth of the product?
System Hypothesis
Turn one product reference into a complete, client-ready photography set.
I proposed a workflow where the studio provides a product reference — a sketch, render or source image — and the system handles the rest of the visual production process.
For each product the system needed to understand the product and its category, retrieve catalogue and preference context, design an appropriate environment, determine realistic placement, choose materials and lighting, generate one representative scene, evaluate it, preserve the approved room and product across views, generate the remaining shot family, and package the outputs for ecommerce use.
Responsibilities were separated across five parts
Interior design agent — plans the environment before anything renders.
Catalogue and preference retrieval — supplies approved and rejected context.
Production harness — owns guardrails, retries, thresholds, routing and publishing.
Eleven stages, the roles that run them, and the one human the system escalates to. The three accented stages — concept eval, visual eval, confidence routing — are what made the economics work.
The first goal was not full automation. It was reliable, client-acceptable output.
What Broke In Production
The first ten products proved the idea. Scaling it exposed everything the model couldn’t be trusted to decide alone.
The initial ten-product pilot looked promising, but client review exposed four failure classes.
Every one of these was individually plausible and collectively unusable. The top row is a judgement problem; the bottom row is a consistency problem.
Client taste
The model understood “luxury” generically, but not this client’s taste.
Product placement
The sculpture could be inside a room without actually belonging there.
Product fidelity
Geometry, scale, materials and bases could drift.
Scene fidelity
Multiple views of “the same room” could become different rooms.
Scaling to a hundred products revealed another problem: catalogue-level repetition. At a hundred products and roughly seven images each, a rejected concept could invalidate an entire seven-image family. With early concept approval around 50%, hundreds of finished images could be produced with no commercial value.
I needed to move judgement earlier than generation cost.
Engineering The Reliability Layer
I stopped treating better prompts as the solution and built control systems around the models.
Preference memory
I converted approved and rejected references into client-preference memory covering lighting, greenery, architecture, furniture density, product dominance, placement, geography and material palette.
A product-aware environment
The room is derived partly from the object standing in it — material, dominant colour, finish and visual weight feed flooring, walls, furniture, metals and lighting temperature.
Approved and rejected references decomposed into six scored signals, which become a structured brief. The room is derived from the product, not from the word “luxury”.
Product fidelity as a hard constraint
Geometry, proportions, base, material, distinctive details and scale stopped being preferences and became part of the quality contract.
Canonical scene state
One approved environment became the source of truth for multiple coordinated views, rather than each shot being an independent generation.
Room drift is not a prompting problem — seven independent generations give you seven rooms. Locking the approved scene turns it into something every later view inherits.
Guardrails
Product geometry and base, alcohol, nudity, Gulf-appropriate context and geography, unnecessary water, catalogue repetition, and unsupported product details.
Product fidelity and the canonical scene are the two things the harness refuses to let drift. Everything in the guardrail column is enforced structurally, not asked for in a prompt.
Evals before cost
A concept eval before rendering, then one proof image scored by a multimodal eval before the remaining six views are produced.
The two evals sit in front of the expensive step. A concept dies for the cost of scoring a brief; a bad set of seven dies for the cost of seven renders.
Confidence-based routing
High-confidence concepts continue automatically; low-confidence cases go to a human. In a typical 500-product example, roughly 450 continued automatically and around 50 required manual review.
Approximate project figures. The routing example is a five-hundred product batch — roughly fifty first images reached a person.
The system was not designed to eliminate humans. It was designed to use them where their judgement produced the most value.
Operating Impact
From a ten-product experiment to a workflow capable of handling thousands.
The project evolved through distinct stages, each one exposing the constraint that defined the next.
~10 products — manual pilot.
~100 products — early automation, roughly 50% approval.
Another ~100 — with prompt and image evals, roughly 75–80% approval.
Human-reviewed first images — near-perfect approval, but too much manual review.
~500 products — confidence-routed, roughly 95% approval with around 90% less client review.
Architecture designed toward ~2,000-product batches.
Before
After
To photograph
Manufacture, crate, transport, install
A sketch or reference image
The environment
Secure a premium location, style it
Planned by the interior design agent
Taste
Argued per image
Preference memory and a structured brief
The seven views
Seven independent generations
Seven views of one locked scene
Rejection
After ~700 images exist
At the concept, before rendering
The client
Reviews after expensive work is done
Routed to, on low confidence only
Compliance
Hoped for in the prompt
Enforced by the harness
~50% → ~95%
Concept approval, early automated to confidence-routed
~90%
Reduction in manual client review
10 → ~2,000
Batch size the architecture was designed toward
Most importantly, the studio no longer needed to physically manufacture and stage every catalogue item purely to create environmental marketing imagery.
The value was not cheaper image generation. It was marketing a much larger catalogue without physical-production cost rising linearly with catalogue size.
On these numbers. They are approximate project figures rather than exported logs, and are stated that way deliberately. They describe the shape of the improvement, not an audited measurement.
What This Taught Me About Deploying AI
The hard part wasn’t getting a model to generate an image. It was building the system that knew when the image was good enough to use.
What the engagement actually covered
LLMs for interpretation, planning and prompt generation.
Multimodal evals for product fidelity, placement, room quality and consistency.
Retrieval and preference memory over approved and rejected references.
Guardrails for hard constraints.
Evals placed before expensive generation.
Canonical scene state for consistency across a shot family.
Confidence-based human-in-the-loop routing.
Orchestration for retries, thresholds, QA and publishing.
The shape of the work was: business problem, prototype, deployment failure, system redesign, client feedback, scaling, operational impact. Each stage was defined by the constraint the previous one exposed.
This was never about finding the perfect model. It was about building enough structure around imperfect models that they could reliably do real client work.
FAQ
The questions this project usually gets.
Why didn’t you just use a better image model and improve the prompts?
That was my first instinct, and it worked up to a point. The problem was that the failures were not only prompt-quality failures — they spanned client taste, product fidelity, scene consistency, catalogue repetition and cost.
A better prompt could improve one image. It could not reliably preserve the same product across seven views, or decide when a bad concept should be rejected before I spent money rendering the whole set.
I eventually treated the model as one component inside a larger production system rather than the system itself.
What was the hardest part to make reliable?
Product and scene fidelity. The images could look beautiful while still being wrong — a $30K–$40K sculpture could come back with a slightly different base, changed proportions or altered material, and the room could drift between angles.
I had to make fidelity a hard constraint and introduce a canonical scene, so every later view inherited the same approved environment instead of rebuilding it each time.
How did you get the AI to understand the client’s taste?
I stopped using vague prompts like “luxury interior” and started treating taste as structured context. I asked the client for approved and rejected references, then broke that feedback into lighting, greenery, architecture, material palette, furniture density, geography and product dominance.
That became preference memory for the interior design agent.
The idea was not to ask the model to magically have taste, but to give it a reusable representation of the client’s taste.
Where did you use AI, and where did you deliberately not use it?
AI was useful for product interpretation, scene planning, prompt generation, multimodal evaluation and visual reasoning.
I did not let the model freely decide hard constraints — product geometry, prohibited content, scene-state rules, or whether an approved room should change. Those belonged in guardrails and deterministic workflow logic.
AI for interpretation, rules for non-negotiables, humans for ambiguous creative judgement.
How did you stop bad generations from becoming expensive?
I moved evaluation earlier in the workflow. At a hundred products and roughly seven renders each, rejecting a bad room after the full set was generated could waste hundreds of finished images.
So I added one eval at the concept stage, then generated a single proof image and ran a multimodal eval on that. Only concepts that passed both gates earned the remaining six renders.
Reject while the mistake is still cheap.
Why did you still need humans if the goal was automation?
Because taste remained the hardest part to fully automate. Reviewing every first image gave excellent quality, but on a 500-product batch that meant 500 client decisions.
So I introduced confidence-based routing: high-confidence concepts continued automatically, lower-confidence ones went to the client. That reduced client review by roughly 90% while keeping approval around 95%.
The goal was not maximum automation. It was the best operating point between AI cost, human attention and quality.
What did the production architecture look like by the end?
A staged pipeline: ingest the product, retrieve catalogue and preference context, plan the environment, run a concept eval, generate one proof, run a visual eval, route by confidence, lock the approved scene, generate the full shot family, run final QA, and package the assets for ecommerce.
The important part was the orchestration around the model — context, memory, evals, retries, thresholds, scene state, escalation and publishing.
What would you change if you rebuilt it today?
I’d invest earlier in observability and eval infrastructure. I’d want better tracking of:
Why concepts fail
Which guardrails trigger most often
How often humans override the confidence score
How approval varies by product category
Where catalogue repetition starts creeping in
I’d also make model swapping easier, so the system could compare generators without changing the surrounding workflow.
Production AI quality depends as much on measurement and routing as it does on model capability.