TL;DR. Generative image models are good at mood and bad at facts. For a manufacturer that is a real problem: one screw too many, a connector in the wrong place, an invented nameplate. ProductLLM is our research project to fix this. Several AI agents plan, check and render product images from the real geometry. Our own ML model learns from synthetic variants to recognise products in photos component by component and to place them exactly. Every image is checked against the CAD before anyone sees it.
The name is a little misleading
ProductLLM is not a single language model. It is an orchestration of several specialised agents and our own model for geometry. The language models do what they are good at: reading, planning, reasoning. The geometry stays in an engine that calculates instead of guessing. The result is an image that looks good and is technically correct.
Three images, one geometry
Our demo shows this with a servo motor with planetary gearbox that we designed ourselves: 34 parts in 7 assemblies. Point cloud, model and scene come from the same geometry, with the same camera and the same focal length. Scroll through slowly.
Point cloud · model · scene
- 01Point cloudsampled from the surface
- 02Modelassemblies and wireframe
- 03Scenelight, material, environment
Agents instead of a prompt
A prompt alone does not know your product. So we split the work. First, the CAD assembly is turned into a visual model that only contains the visible shell. Your engineering know-how stays with you. A planner agent reads the assembly, the bill of materials and the communication goal and plans the scene: camera, lighting, exploded view, highlighting. A checker agent tests the plan against brand rules and mandatory components before the engine renders.
The chain, step by step
Only then does generative AI step in, and only for the surroundings: light, material, background, an in-use scene. The refinement is bound to the depth and edges of the rendered image. At the end, every image is laid over the CAD model and checked component by component. If a part is missing or misplaced, the image goes back.
The other way round: from photo back to geometry
The same engine can reverse the direction. It generates thousands of variants of a product, in every pose and every light, with exact labels for component, pose and depth. These labels cost nothing, because the engine knows them while rendering anyway. From them, our own model learns to trace real photos back to components and 3D geometry, and can place a product exactly in a real environment.
Boxes straight from the geometry
This leads to three applications we are working through with manufacturers:
What it is good for
- Scenes that hold upProduct and in-use images for marketing and sales, technically correct down to the screw.
- Spare part by photoA customer sends a photo, and the model tells you which part is meant.
- Digitise legacy productsTrace products without CAD back to geometry from photos.
Where we stand
ProductLLM is a research project. The demo is in place, the architecture is designed and written up. We are tackling the scientific part, above all reliable checkers and the model for real photos, together with a university AI institute. Whatever is ready goes into Brand Studio.
AI may build the stage. The product stays exactly as it is.
What you should do now
- Check AI product images component by component. Count screws, connectors and labels. Whatever you find is the benchmark for any solution.
- Treat CAD data as an asset. For images, the visible shell is enough. Your engineering know-how never has to leave the building.
- Collect your cases. Where do you need product images in variants, where do spare-part requests arrive as photos, which products have no CAD?
If you would like to see ProductLLM on one of your own products, just get in touch.
