I am Scott Davidson and work at Robert McNeel & Associates. I have been experimenting with the
AI
panel in the Rhino 9, especially how to model with it. Here is an architectural modeling workflow I found quite interesting. At the moment, it seems that ChatGPT Astra is the best AI model for 3D models, so I have linked it to the
AI
panel. The AI panel in Rhino automatically gives the AI some additional access to Rhino. There are many different ways to approach modeling with AI, but this one has gotten me the fastest results yet.
What You’ll Need
- Rhino 9 for Windows or Mac.
- An MCP-capable AI client that can connect to a local MCP server. In this guide, I linked ChatGPT Astra to the AI panel, since at the moment it seems to be the strongest model for 3D work.
- A subscription to that AI client. Everything here was done on the standard $20 Astra plan.
Verifying your Connection
In this section, we will make sure we have the panel open and docked in a suitable way to start our conversation. We will verify that the client is connected. The AI panel automatically gives the AI some additional access to Rhino, so there is nothing else to set up.
- If not, open the
AI Settings
to verify which Agent is available and eventually select a Model:
- If it is still not connected, run
/logoutfollowed by/loginby typing it directly in the AI Panel’s chat box. This should force the connection.
Start with an Image
I started with a prompt to create an image of the stadium:
Create an image of a modern football (soccer) stadium. The stadium has a sweeping oval steel roof supported by elegant angled columns, a glowing glass central facade, and patterned metallic exterior panels. Warm interior light spills onto a broad polished concrete plaza as crowds of fans approach the entrances. Cinematic architectural photography, wide eye-level view, cool twilight sky, subtle reflections, crisp contemporary materials.
It came up with a series of images like this one:
The input for the design ideas could be quite varied. Hand sketches, context images, program requirements, and other models could all be used to influence this initial design. What I have found is that working out design ideas at this stage is very fast.
One Image Isn’t Enough
However, in my experience, using just a single rendered image to model does not give great results:
Build a Set of Reference Images
Things improve when I create a series of images that look at the design from multiple angles. Here are the prompts and the results:
I need a picture of a scale architectural study model of this (image above) using museum board as the medium.
Can I see a view from above the model?
An image from inside the stadium?
The white model is a series of images that themselves are AI generated. They were inspired by the original design idea set down with the very first rendered image. It was these white style images that helped inform the better 3D model. Without them the 3D was very poor.
Model from the Image Set
Create a 3D model of this sports arena using the images attached.
As you can see, the result was much more detailed and closer to the original design image at the start of this process. To me, this is the key takeaway so far: adding more images earlier in the process moves the 3D model further along, faster. Then use all the images created to inform the model.
Of course, we could now continue to refine the 3D model with text prompts or additional images if needed. Engineering data, such as the required layout of the bowl and seating or circulation requirements, could also be brought in to influence and change the 3D model through the AI panel.
What’s in the Model?
The model itself is properly layered, with materials already considered for rendering. Extra BIM data and block instances can also be created.
I did not touch the model directly or mention a geometry type, and it created 818 polysurfaces and 6481 meshes.
I can tell AI that I want NURBS objects instead and it will work through that where possible. That is the interesting part of this. AI is making many inferences as to what the object is and how it should be modeled based on many clues like if it is a panel, or a roof, or a floor. These are things I have not told it, but useful when working with the AI.
Render it from the AI Panel
Then, what’s fascinating is using the AI panel to render this model:
Create a photorealistic rendering of this sports stadium from the perspective view:
It’s a very quick way to design a model and then get back an image of it that closely parallels your design images, all in a short amount of time.
Adding Detail to the Bowl
Here is an example of adding detail to the design. The ability for it to fill in the blanks is both amazing and frustrating.
Here is what was first modeled for the interior:
The original image of the interior has some detail to it. It did not translate to the bowl initially, but something like this started to work:
Create the aisles, tunnels and field walls shown in this image into this 3d model of a stadium
It did read the existing model before proceeding:
The model already has radial stair aisles in both seating tiers. I'll add the aisle parapets, rectangular tunnel entrances, and segmented field walls from the reference, aligned with those existing aisles.
I'm adding eight tunnel entrances per tier, with open passages beneath the seating and matching breaks in the field wall. The new walls will use the model's cream finish and sit on separate layers for easy editing.
Added to the stadium model: Parapets along both tiers' existing stair aisles, 16 open tunnel entrances with side walls, roofs, and floors. Segmented field walls with aligned entrances and ramps. New geometry is on three separate interior layers. All 296 added objects passed geometry validation.
And it gives me this:
But of course that is not what I want, so I gave it more detail:
The lower tunnels need to be 5 rows back from the field. and enter on a pathway that travels all the way around the arena. The upper tunnels also need to be 5 rows back entering onto a path. And the upper tunnels should be offset from the lower tunnels so they alternate.
AI gets back to me:
I'll move both sets of tunnel entrances five rows back, add a continuous circulation path at each tier, and stagger the upper entrances halfway between the lower ones.
I'll keep five seating rows in front of each path and make both paths continuous around the bowl. The upper tunnels will rotate by 22.5°, placing them between the lower tunnels. I'll also restore the seating where the original front-row entrances were cut.
Updated both tiers with five seating rows in front of continuous arena paths. All tunnels now open onto those paths, and the upper tunnels are offset 22.5° to alternate with the lower ones. Restored the former entrance cuts and adjusted the stair connections. Geometry validation passed.
Here is the result:
Roof, Floors and Beyond
The same sort of process can be run for the roof structure and other areas of detail:
The floor spaces on the outside were done the same way:
And we can end up here:
What Did it Cost?
I do not have all the token counts. But here is what I do have:
Creating the tunnels and moving them around - 626k tokens
Creating the roof extension, creating the two exterior floors, and a rendering - 772k tokens
Token cost specifically is hard to calculate. This was done within the standard Astra $20 subscription. Although, if I really would work out more details, which I would have to do if it were a real project, I suspect the $100 subscription tier would be required.
Take Aways
An experiment that would be interesting is to add many more detailed images before modeling. Then see how far it might go with more detail at each level. Like I covered above, spending more time making images first does really benefit the model later. I just do not know what the limit is.
There are really so many levels of interaction and design development best practices that still need to be worked out. We are all learning on the fly.
Join the conversation on the original Discourse thread.
