Modelling a Rose in Autodesk Fusion with Claude Fable 5.2

Back in December 2009 I wrote two posts on this blog (yes, the blog is really that old): Complex Surfacing in SolidWorks and Surfacing in Alibre Design. In both of them I modelled a rose. I had started playing with surfacing in Mechanical CAD software to understand the limits of the software, and the limits of myself.

Lately I have been reading about, and watching videos of, people using AI to model things in CAD. Almost all of it is prismatic: the brackets, housings and plates that are common in mechanical design. AI has clearly been trained well on such designs. I wanted to know how well it has been trained to model organic objects in MCAD software like Autodesk Fusion. So, 17 years later, I went back to the rose.

I gave Claude Fable 5.2 a phone photo of a red rose and asked it to model the flower in Autodesk Fusion so that a render would match the photo as closely as possible.

A random picture of a rose I got from the web

Claude was connected to Fusion through Fusion’s MCP server. It wrote Python against the Fusion API, ran it in my session, checked its own screenshots and test renders, and corrected itself. I gave a handful of short instructions and never touched the mouse. This is what I got.

The final render, straight out of Fusion’s local renderer with no retouching

How long it took

From creating the new Fusion document to the first full-resolution render took about 11 minutes. The scripts that built the geometry ran in about 13 seconds. Most of the rest went on test renders to get the camera and exposure right.

Three further rounds of refinement followed: petal rims, hooked thorns, translucent petals and more visible dew. From the new document to the final render shown here took about 46 minutes, including the time I spent looking at each result.

Inside the model

Nothing was sketched by hand. Claude’s scripts computed points, fitted splines through them and called ordinary Fusion features, so every loft, thicken and pipe sits in the timeline.

• The petals are solids, not surfaces. Each is a lofted surface thickened to less than a tenth of a millimetre, so the edges stay sharp and the translucent petal material still has a volume to work with.
• The 26 petals spiral. They are spaced at the golden angle, as in many real plants, and no two are identical.
• The leaf veins and serrated edges are geometry. They are modelled into the surface of each of the 13 leaflets. There are no textures or bump maps.
• The small things are modelled too. The thorns are hooked, and 152 dew drops sit on the petals and leaves.

Screenshots

Where it falls short

This is a convincing, clean CG rose, not something you would mistake for the photo. The petals and leaves are a single flat colour, neighbouring petals pass through each other near the base, and the dew drops are perfect spheres.

But Damn! What a time to be alive! For someone who actually understands how much time and effort it takes to model intricate organic chapes in Mechanical CAD software, this is absolutley amazing. Truth be told, I’m lost for words. And the best part is this is the worst that AI will be. It will only get better from here.

What this means for CAD modelers

The part of the job that Claude did here is the part many of us think of as the job: turning an idea into features, one click at a time. It did that for an organic shape, in a mechanical CAD system, in minutes. If a modeler’s value lies mainly in knowing where the commands are, that value is shrinking.

What Claude did not do is decide that a rose was worth modelling, judge whether the result was good enough, or know what the model was for. It made choices that suit a picture and would fail anywhere else. Petals pass through each other, and the leaves have no thickness. A modeler who understands design intent, manufacturing and fit would catch those in seconds. That knowledge is not going away. It becomes the thing that matters.

So I do not think CAD modelers will be sidelined by the technology. I think some will be sidelined by other modelers who use it. The work shifts from building geometry to specifying it and checking it. That is a real change in skills, because describing what you want precisely, and spotting what is wrong in something you did not build, are harder than they sound.

Do modelers have a choice? As individuals, for a while, yes. Plenty of work is too specialised, too regulated or too poorly documented for an AI to do well today. But the larger choice is not theirs. Autodesk ships the connector that made this experiment possible. When the CAD vendor builds the door, employers and clients will expect people to walk through it. We have been here before, with the move from the drawing board to 2D CAD and from 2D to 3D. Nobody was forced to switch. They simply found that the work had moved.

One rose is one experiment, and a rendered flower is not a production part. But I started surfacing in 2009 to find the limits of the software and of myself. This time the limit I ran into was neither. It was how well I could say what I wanted.