We Built an AI Photography Pipeline. Here's What We Learned.

By Ernests, CEO of LuceneX

The Problem: A Thousand Products Nobody Could Photograph

Our client manufactures furniture, solid, well-built pieces sold through e-commerce. They had a problem that sounds simple until you think it through: they needed professional photography for over a thousand products, and traditional photography was economically impossible.

Picture the logistics. Their products sit disassembled across warehouse pallets, with components weighing 30 to 60 kilograms each. To photograph one assembled piece, you need workers to locate the right parts across multiple pallets, transport them, assemble the product, stage it for photography, shoot it from multiple angles, then disassemble and return everything. Now multiply that by a thousand products, each needing four angles minimum, eight variations for modular components, and lifestyle shots showing the furniture in realistic home environments.

The math gets brutal fast. We estimated traditional photography would run into tens of thousands of euros: warehouse labor, assembly time, professional photographers, studio space, post-processing. For a furniture manufacturer with finite marketing budget, that's not a line item. That's a barrier to market.

They needed a creative solution. That's where we came in.

Why AI: The Middle Path

We considered the obvious alternatives first.

Option A: Reduce scope. Photograph only bestsellers, skip the variations. The client rejected this. They wanted their full catalog represented properly.

Option B: Pure 3D renders. I have basic 3D modeling skills, so we could theoretically render everything digitally. But here's the problem: nobody buys furniture from a video game screenshot. When a customer opens a product page, they need to imagine that piece in their home, as part of their daily life. A synthetic-looking 3D render breaks that imagination immediately. It's the uncanny valley of e-commerce.

Option C: AI-generated photography. This became our answer, but not in the way we first imagined.

The client had one asset we could work with: photographs of individual components. My initial thought was simple. Feed the AI pictures of various parts and hardware, then ask it to assemble them into a complete product image.

That failed almost immediately.

The Build: 3D Models as Skeleton, Photos as Skin

Here's what we learned: generative AI models like Google's Gemini don't understand physical assembly. They don't know that this component attaches here at this angle with these proportions. Without that structural context, the AI hallucinates. Wrong angles, impossible joints, proportions that don't match reality.

A colleague suggested the breakthrough: what if we created 3D models not as final outputs, but as references for the AI?

So that's what we built. I created accurate 3D replicas of each product in Shapr3D and Blender with proper dimensions, correct assembly, one-to-one copies of the real furniture. Then we rendered these models from multiple angles.

The pipeline became:

3D model provides structure and assembly context

Component photographs provide real textures and materials

Detailed prompts describe the desired output

Gemini API generates photorealistic final image

The AI wasn't assembling from scratch anymore. It was essentially "skinning" a known structure with real-world textures. The 3D render told it how things fit together. The photos told it what those things actually look like.

The Pain: Two Months of Sub-60% Success

I won't pretend this worked smoothly. For roughly two to three months, the success rate sat below 60%. More than half our generations were unusable: wrong angles, incorrect textures, hallucinated details.

I started with maybe 40 reference renders. That became 100. Then 200. Then over 1000. Each iteration taught me something about what the AI needed to see.

The prompt engineering was equally brutal. I went through five major versions of our prompts, each with subversions, each subversion with its own iterations. I generated somewhere around 5000 images before reaching production quality.

There were moments I genuinely thought I couldn't continue. The work becomes monotone, exhausting. At some point you start dissociating, convincing yourself an image looks real when part of you knows it doesn't. That's when having another pair of eyes became critical. Someone not involved in generation who could say "this doesn't look right" when I'd lost perspective.

Google's AI Studio, their native Gemini interface, was another obstacle. It's restrictive and counterintuitive for batch processing. I eventually built a custom system using Python and JSON-structured prompts that let us process multiple images per product, multiple products per batch. The boring infrastructure work that made everything else possible.

Two Problems: Clean Shots and Lifestyle Images

The project split into two distinct challenges.

Phase one: clean product images. These show the assembled furniture against neutral backgrounds, what you'd see as the primary product photo. Once we stabilized the prompts and 3D renders, this became relatively manageable. I could compare outputs against real products I'd physically examined, verify proportions, and batch process with reasonable consistency.

Phase two: lifestyle context images. This is where AI still struggles hard.

These are the aspirational shots: the furniture in a warm, lived-in space with people interacting naturally. The images that help customers imagine the product in their own lives.

AI-generated humans are problematic. We'd get images with nearly identical people in the same room, two figures who looked like clones standing across from each other. Uncanny and unusable. Proportions would drift unrealistic. Products would appear in environments where they'd never logically exist.

The prompt specificity required for lifestyle images is much higher, and much easier to break. You need realistic environments without being overly restrictive. You need props that make scenes feel lived-in without looking staged. Every additional element is another variable that can go wrong.

We got it working, but there are still skeletons buried underneath.

The Results: Significant Savings vs. Tens of Thousands

Here's the number that matters: I spent approximately a few hundred euros on API costs to generate over 5000 images.

The traditional alternative (warehouse labor, assembly, professional photography, editing, repeated across a thousand products with multiple angles and variations) would have cost tens of thousands of euros. That's our rough estimate, and even being conservative, the cost reduction approaches 80%.

Beyond cost, there's time. What would have taken months of coordinated warehouse and photography work compressed into a pipeline we could run iteratively, improving as we learned.

The client now has a complete product catalog with professional-quality imagery. Their Shopify store launched with full visual coverage, something that simply wasn't economically viable before.

What We'd Do Differently

If I started this project tomorrow, with the scar tissue I've developed:

Start with 3D models immediately. Don't waste time hoping the AI can figure out assembly from component photos alone. It can't.

Leave no room for ambiguity. Be specific about expected outcomes. Vague prompts produce hallucinated results.

Skip restrictive interfaces. Build batch processing infrastructure from day one. The time investment pays off immediately.

Get external eyes early. You will lose objectivity. Plan for it.

The Responsibility That Comes With This

This project taught us something beyond technical skills: AI-generated product photography carries ethical weight.

If you can generate professional-grade images from 3D renders and component photos, you can also mask defects. You can misrepresent quality. You can show customers something that doesn't match what arrives at their door.

That's a thin line, and businesses need to walk it carefully.

Used responsibly, this technology democratizes e-commerce. It eliminates the massive starting capital that kept small manufacturers from competing. You need less infrastructure, less specialized knowledge, fewer tools to bring products to market.

But the responsibility remains: make sure what the AI generates matches reality. Your customers deserve that. Your business depends on it.

LuceneX builds AI-powered solutions for businesses facing impossible tradeoffs. This project combined 3D modeling, custom pipeline development, and Google's Gemini API to solve a photography problem that traditional approaches couldn't touch economically. [Contact us](https://lucenex.lv) to discuss your challenge.