Hi,
Anyone know if this new kid on the block, ChatGPT can be fed a set of images all same focal length, and generate accurate CAD model from them ?
Cheers
Steve
Yes and No
The LLM needs a conection to the software to be able to do it, that is what people call MPC (or MCP don’t quote me on that) and it usually have to have a skill set so the LLM Model know what to do and what it can do.
There is an official McNeel repository and plugin for this, the plugin is called RhinoMCP Platform and the github has some information on how to setup using claude code.
There are other MCP solutions out there in the plugin savana, but for now I would be careful with using it since it has a direct connection with your files.
One of the last updated ones that at least worries about a disclaimer of the geometry is Aurox.
https://www.food4rhino.com/en/app/aurox
But again, use this kind of stuff from third parties with care, since you usually need an API and that can get costly very quickly if you API key is leaked or something similar.
Thanks for replying, Sounds a bit tricky and dicey.
Has anyone reading this got a decent result from ChatGPT making a Rhino model ?
I am 2 days processing and tweaking in Adobe Camera Raw 665 photos of a trailer bowser WW2.
Just had the thought of this ChatGPT making a mesh then making a model from it, saving me weeks of work !
Maybe it even could take the photos and seek out shadow detail, save me adjusting each photo.
Steve
If you do use an AI Agent to model from an image, you’d likely just give it one image and it’ll figure out the rest if it knows, or guess if it doesn’t. Then slowly feed in more images and get it to iterate if possible.
In my experience using the McNeel MCP Server with Claude it turns the image into a python script and creates a single surface or mesh of the object. It would be cooler if it could walk through and model it using commands but I’ll need to experiment more.
If you are able to post one of the photos here I would be curious to give it a go and report back as I’ve only ever tried simpler things.
What you described sounds more like a COLMAP - Structure-from-Motion and Multi-View Stereo — Johannes Schönberger problem than a LLM problem.
These MCP connection tools don’t really do a goob job to recreate something, they do generate some stuff but its usually a new stuff, not really a recreation of true dimensions.
This is a reverse engineering application, beginning with photos is a step to generating a mesh, and from the mesh to a CAD model.
As @cdordoni suggests, obtain a high quality mesh with photogrammetry software, then reverse engineer in CAD. Time consuming, but works very well.
I’ve seen this done with meshes but not NURBS like in Rhino.
RobinAi, HunyuanAI, TripoAi
Hitem3D
Steve has used Metashape photogrammetry in the past, and my guess is the 665 photos he is processing were intended for use in Metashape. It looks like he is hoping for a quick way to a model using AI.
Apparently Steve is adjusting the contrast, brightness, etc in each photo individually which is taking a long time. My experience with Metashape is different. I batch process the photos (if I don’t use the jpegs from the camera) which usually takes less than an hour.
Photogrammetry is AI, it is not language model AI, but it is AI. You just can’t type it what you want in natural language.
Most mesh generator AI aim for “good enough” since most of the use of these things is for 3d printing or “artistic models” not for proper engineering.
For now, at least as far as I know, there is no model that can beat the good old photogranommetry. Apart from the obvious 3d scanner path, that is another subject.
It’s all “machine learning” algorithms that have been around for 70 years, calling chatbots AI is an insult to the concept of actual “AI.”
Photogrammetry is not AI (though there may some photogrammetry software which claims to be AI enhanced or similar). Photogrammetry is deterministic using defined algorthms. There is not machine learning, neural networks, or similar.
I remember when it was called “machine learning”
then media got hold of it and blam!
Everything is called AI now…
If it can not hallucinate enough, it should not be AI
Photogrammetry is quite procedural and parametric, so more on the side of the algoritmic world than the nowadays here and everything fuzzy-touchy-guessy AI
(One of) the less deterministic part of photogrammetry may be the initial surface fitting, commonly treated via the damped least squares algo
Funnily, this algo shares a common method with today AI eg. gradient descent
But the way it is applied varies drastically, and no surprise AI introduces a loss of robustness, quick convergence, etc.
If you want to go further, you can ask to grok or gemini to
compare gradient descent in damped least-squares vs in nn learning
You can even verify this whole post on grok or gemini o:
Try Meshy it is amazing
I found Meshy incredibly poor. You cannot upload a range of photos that captures the product from all angles, and all you get is a polygon mesh, no surface or solid model that you can then use in the industrial design and downstream product development process.
It only works with symmetrical objects, and of course you still have to scale the object afterward.
In my case, it’s basically useless, because I work in the footwear industry. If you look at the topic, I tried providing multiple photos of the same heel, but the AI interpreted them as separate objects, so it gave me multiple meshes back.
I emailed the developers explaining that, for things like soles, wedges, and heels, we need the option to provide multiple photos of the same object from different angles. Otherwise, it will never be even remotely usable for this type of work.
In this example, I generated this kawaii cat with Midjourney and then fed it into the open-source AI.
Instead, in this case, I took an image from Google and fed it into the AI.
I really like how the blown-out highlights combined with the white background caused the 3D model to become concave or completely missing geometry in those areas. The fact that such a model can be generated is impressive, but these moments of exposing the AI’s intelligence are always charming.
That’s why I like photogrammetry, if the client pays for it. At least you have a highly accurate mesh to then start reverse engineering.