Hey everyone, I wanted to open a discussion about how you’re integrating AI into your workflows.
Imaging isn’t my main focus, though I’m curious to hear what the state of the art looks like on that front for you. On the modelling assistance side, I tried Raven, but it didn’t win me over. Explaining to Raven what I need takes roughly the same effort as just building the script myself. On top of that, I often don’t know exactly what I want until I start building it, and that serendipitous discovery is part of the process.
Where I do find myself comfortable is chatting with Claude or Mistral for anything Python-related. When I have a clear problem in mind and I know Python can solve it better or faster, I delegate the code writing to them, and they do it significantly better than I would.
That said, the more I rely on them for this, the less I’m actually learning Python. It’s a trade-off I haven’t quite figured out yet.
I use Mistral, Gemini etcetera most of the time for finding views, I did not think of, or just forget. Example:
I want to to combine two curves to a surface. How?
AI comes up with all components I could use: Loft, Ruled Surface, etcetera.
I use what I need. (and here the help stops).
Or
I start explaining to AI that the given solution is impossible. And my arguing sharpens my thinking (not that of AI, what already forgets what it wrote in the middle of the conversation)
Regards, Eef
I’ve been using Cursor with quite a lot of success in coding projects, combined with ChatGPT and Claude.
It is useful to first work on the design, iterate over that until you have a document that is detailed. Make sure you understand the document and it is aligned with your goals. Then based on the design generate an implementation plan. Have TDD also be a big part of design and plan. Review both design and implementation plans often until it all makes absolute sense.
After that having these tools generate code gives often a useful result. It becomes a matter of doing your design and planning work diligently. Essentially building a very, very specific prompt.
Obviously it helps if you know what you want and how you would implement it. That ensures you actually understand the code being generated and allows for corrective action whenever needed.
Any broad stroke prompting is going to give you Whatever, but being precise and clear will yield much better results.
Currently, what AI does is reproduce a render that maybe 6 times out of 10 looks somewhat similar to the modeled object, with a result that can be satisfactory, but never exactly how I would really want it. The company that, in my opinion, is getting closer is Vizcom, but for me we are still in a phase where AI gives back images that work well for concept design, not for the final industrial product. Vizcom AI YouTube Channel
The AIs that instead generate a 3D result from images are not even remotely close to what is actually needed. They never manage to reproduce an exact geometry based on the reference images you provide. They mostly work with simple geometries and designs that have a certain symmetry. For example, in the footwear industry they almost never manage to produce a real, usable shape. Sparc3D on Hugging Face
For some people this may seem like a good result, but not for me, look here: Example video showing AI 3D defects
There are many defects in the result it gives back. It is fine if you still consider it as a concept or a starting point.
The idea that AI can truly help you with modeling, at least in the way it is being promoted today, is pure fantasy to me. If you think it can really help with industrial design itself, it is impossible. AI exists inside a computer, not outside of it like human beings. It cannot speak with the client, so it cannot truly understand or interpret the client’s needs. If a client gives me specific directions to follow, asking to modify certain areas of the design, AI is not capable of handling that without completely altering what is actually required.
By the time it even starts understanding what is needed, I have often already finished the model myself.
However, I do think AI could have real usefulness in specific commands. For example, imagine having to build a G2 surface. If AI could automatically test different surface generation methods and give me the best possible result — with proper G2 continuity, the correct degree, and the lowest possible number of spans — then yes, that would be fantastic.
Another smart use of AI that is actually becoming interesting is the automation of CNC commands.
I am not. For 3D surface modelling of products, it is entirely useless. I hope that the AI enabled G1, G2, and G3 surface continuity matching tools Dassault and Autodesk are working on will come to fruition in my lifetime.