Well, to be honest, you can have an AI assistant do the workflow maintanence.
But we do have tools to better connect Comfy to Rhino already.
It is basically Grasshopper, but for image/video generation.
I looked at Comfy UI regarding the use of meshes of finished design concepts or finished designs, but, as mentioned above, it’s a very tedious process, and there are all kinds of issues reading the OBJ and MTL files from various 3D software via Load3D, setting up a product photography camera, generating the various depth maps for the desired viewing angles via ControlNet, etc. The Alman Tools example shows a very vaguely defined building design that in the adjacent rendering looks totally different. I don’t find that example convincing at all, and I would rather not use Grasshopper to design a novel portable defibrillator. Industrial designers need to focus their valuable time on strategy, innovation, and then designing, not noodling with Grasshopper and Comfy UI nodes ; )
Vizcom faithfully does what’s needed, in no time, at little cost.
That looks rather tedious and the result is rather cliché, like a typical Keyshot rendering. Also where do you load and specify all the textures and colours for the product’s CMF? What do you do with the rim, tyre or windscreen wiper data from another CAD system?
It’s pretty obvious why Vizcom has taken the lead.
There’s a prompt where I specified how I wanted the materials, including the environment. It’s located in the program’s user interface. I specified that I wanted the wheels included in my model.
Where do you load the textures though? Metallic car paint needs flakes and a clear coat. Seats need the supplier’s leathers or textiles. How do you put the different types of stitching on the various parts of a seat? Three different views plus close-ups without AI hallucinations? I find that thing very unconvincing and impractical for industrial and automotive design.
I completely agree with your point. I also take a very conservative approach to using AI images when dealing with clients, especially in contractual relationships or commercial projects where money is involved. Clients need to see an exact, true-to-life representation of the final product, not a generalized guess.
However, we haven’t completely given up on AI. Instead, we are trying to bridge the gap by incorporating elements of physically-based rendering (PBR) techniques, similar to what you would find in software like KeyShot. This is exactly why we prefer an adjustable, canvas-based AI that acts as a middle ground between pure generation and precise control.
Ultimately, my team and I aim to develop a fully controllable AI rendering workflow—one where you can input or select specific materials and colors just like in KeyShot. Even if we don’t build it first, I believe a solution like this will hit the market very soon to solve the exact practicality issues you mentioned.
Hi @Lagom
If you haven’t already, take a look at Figurement - ex-Luxion (makers of Keyshot) people building a AI-assisted render engine. I haven’t tested it much, but shows potential.
-Jakob
@Normand Interesting, I took a look at their videos and it looks promising. However, it’s a credit-based system like many AI services, which probably won’t sit well with everyone—some people are fine with it, others aren’t. I’ve always seen mixed opinions on this, but that’s to be expected.
What I’m wondering is whether the files you import into their system get downloaded and stored on their side, because that could be a bit of a concern in terms of privacy and confidentiality.
Also our main concern! We tend to stay away from anything cloud based. I just came across it on Insta, and have made a (private) account to test it out, just using whatever random file I have laying around. I’ll get around to it eventually 
-Jakob
What I liked about the BYRhinoGen AI I tested is that I don’t have to upload the file anywhere, and the renders are saved locally. Of course, there’s always that doubt, but it’s already something compared to uploading a file somewhere unknown. If they actually maintain data privacy, for now it works for me, because it doesn’t reinterpret the geometry of my model — it renders exactly what I want based on my prompt.
At this point, woulnd’t it be better to just train the AI on renders ? Like, you have the rendered fire configurations and the render result and use it as data to fine tune the AI based on the configurations and render. Generate the image and compare with the render and adjust and so on.
After that you train a assistant to apply the configurations of the render, so we remove that time intensive part as well. So when the prompt is given, the AI agent apply the materials as if it would be rendered regurlarly with the cycles engine, and that information is then fed into the generating AI so the final result is more correct to the objective.
Similar in how they are training AI with CFD and FEM. they have the entire setup of the simulation and then use the AI to predict the result to show just in a few seconds, this way the engineer can iterate the job quicker and at the end do a full on simulation to have the proper validated results.
If byRhinoGen were an AI we developed from scratch, we could train it on various datasets. Unfortunately, we are currently leveraging the Gemini API.
My idea is this: when a user selects and inputs a material, both the image of that material and the text prompt are sent together. I believe this will allow for much more precise control.
However, we are still figuring out how to optimize the real-time viewport display and the data transfer to the AI.
It’s something I’ve wondered about for a while. With all the cloud rendering that’s been done over the years, I can’t help but wonder if some of that data has already been used to train AI models. In the end, when you accept the terms of service, you’re often also agreeing to let your data be used to improve the service or for other specified purposes. Let’s be honest—almost nobody actually reads that huge wall of legal text at the beginning. Most of us just scroll to the bottom and click “Accept.” So I wouldn’t be surprised if some cloud rendering services have already used that data to develop or train AI models, where their terms and conditions allowed it.
By the way, it slightly reinterpreted some parts compared to my original model. At the beginning, I didn’t see it as something to present to a client, but just as a way to play around with it and see the result without a critical eye on everything. It still interpreted a couple of details differently. Probably, with a better-written prompt, this can be avoided, but honestly I can’t say the render represents my model 100%. If I have to judge it critically, for now I still rely on traditional rendering. That said, playing around with this AI is also fun.
I am pretty sure this is something impossible to rule out 100% without some fine-tunning
Research has shown that that is one reason why most AI generated product or architecture scenes look so similarly boring, one lifeless pixel sludge after the other. Most of the training was and still is done with the huge amount of cliché photos from about 2010 - until today, and that, in turn, encompasses ever more AI slop, so it’s a self-regressing system ; )
Even via the API, you can apparently not stop gemini-3-pro-image-preview “Nano Banana Pro” from rotating logos or product graphics, if they are facing “the wrong way” (according to what the AI model “thinks”).
Hey @Lagom Runchat creator here. We’re taking a slightly different approach to viscom / gemini etc on these kinds of creative tasks in that we’re trying to make it possible for agents to iterate on images and judge them against a brief so you aren’t having to do the kind of work outlined in this thread.
This works especially well if you connect claude / chatgpt and let claude drive the agent in a loop - you can then see all of its iteration and experimentation on the canvas and pick outcomes you like best.
Let me know if you want to try it and I can shout you some credits (or try using the free/very cheap models).