Why We Removed AI Renders

We built AI photorealistic rendering into Manufacts. It worked. The images were gorgeous. Then we deleted it, on purpose, and this post is about why — because the reasoning applies to a lot more than our product.

What we built

The pipeline was the obvious one, competently executed: take the depth buffer from our live 3D scene of the configured home, run depth-ControlNet img2img over it, and get back a photorealistic image whose geometry matched the configuration. Cabinets where the cabinets are, windows where the windows are, warm light, believable materials. In a demo, it produced audible reactions. Buyers don’t gasp at wireframes; they gasp at photographs of a home that doesn’t exist yet.

By every conventional product instinct — and every 2026 roadmap template — this was a feature you keep. AI-generated product imagery is table stakes in sales software now. We ripped it out anyway.

What the renders actually did

The problem is in the phrase “geometry matched the configuration.” Geometry mostly matched. Everything else was the model’s opinion.

The diffusion model doesn’t know your catalog. It knows what kitchens look like on the internet. So it would render a beautiful faucet — not the faucet in the order. Countertops came back with a subtle marbling the customer never picked. Cabinet hardware appeared that no rule ever included. Once, memorably, it added a window. A nice window. A window with no header in the BOM, no opening in the shearwall math, no line in the order.

Every image was plausible and every image was wrong somewhere, and — this is the killer — wrong in ways no one could enumerate. With a bug, you can write a test. With a generative render, the errors are sampled fresh each time from the model’s priors. We could not put a bound on the drift between the picture the buyer fell in love with and the configuration the factory would build. We could only make the drift prettier.

A render that drifts from the ordered configuration is a beautiful lie. And in our product, it would have been a beautiful lie positioned as the moment of decision — the image on screen when the customer says yes. The most expensive purchase most families ever make, anchored to an artifact we couldn’t verify.

The standard we actually hold

Here’s the principle that made the decision for us: anything the buyer sees must be derived from the same model that resolves the BOM.

The price is derived from the model — it’s the recursive cost roll-up of the resolved BOM. The option availability is derived from the model — the rule engine greys out conflicts with a reason attached. The order’s explanation, years later, is derived from the model — frozen rule sets re-derive why every rule fired. Everything in the product is a projection of one source of truth, which is what lets us verify it, version it, and stand behind it.

The AI render was the one artifact in the entire system that wasn’t. It was correlated with the truth, seeded by the truth, and not accountable to it. It was the translation layer — the thing our whole company exists to delete — sneaking back in through the demo.

So the walkthrough customers get is Three.js, rendered live from the actual configuration. Choose the other countertop and the countertop changes, because the walkthrough and the BOM read the same data. It is less photorealistic than the diffusion output. It is also true: every object in that scene corresponds to a resolved line item, and nothing appears that the order doesn’t contain. We describe it as a digital twin and we mean the words literally — same model, second view.

Honesty beat prettiness. It wasn’t close.

The contrarian part

Now the broader claim, since the industry is sprinting the other way: AI-generated product imagery in sales is a defect being marketed as a feature.

Not because the images are bad — because they’re good. A bad fake gets caught. A good fake gets signed. When a generative model produces your sales imagery, you have inserted an unauditable artist between your product truth and your customer’s expectations, at precisely the point in the funnel where expectations become contracts. Every drifted detail is a small promise your factory never made. In low-stakes e-commerce maybe you absorb that as returns. In configure-to-order manufacturing, the “return” is a built home with the wrong countertop and a customer holding a screenshot.

The usual defense is “it’s just illustrative, there’s a disclaimer.” Watch a buyer for five minutes. Nobody experiences a photorealistic image of their configured home as illustrative. The realism is the message. If the image weren’t being read as truth, it wouldn’t be selling — that’s the entire mechanism. You cannot simultaneously claim the render drives conversion and claim nobody relies on it.

To be precise about scope: we’re not anti-AI — an AI co-pilot drafts BOMs and floor plans in our admin portal today. The difference is where the human verification sits. A drafted BOM lands in front of an engineer whose job is to check it before release; every line is inspectable against the schema and the rules. A render lands in front of a buyer at the moment of maximum trust with no verification step at all. Generative output belongs upstream of review, not downstream of it.

The test we’d offer anyone

If you’re evaluating configurable-product software and the demo shows you a breathtaking image, ask one question: is this picture derived from the same data structure that produces the bill of materials?

If yes, you’re looking at a digital twin. If no, you’re looking at concept art with your logo on it — and somewhere in your future is a conversation that starts with a customer pointing at a picture.

We chose the boring render. Our customers are buying a home, not an image of one. The picture they configure is the home they get, down to the line item — and we can prove it, which is the only kind of beautiful we’re interested in shipping.