AI can generate plenty of images that look like interesting materials. That doesn't automatically make it useful for CMF. Automotive CMF has constraints that an image generator doesn't understand. A material may need to be manufactured at scale, meet durability requirements, behave consistently under different lighting, and connect to real supplier and sustainability data.
The part we're interested in is much narrower. We want AI to make material exploration faster without asking it to replace the rest of the CMF process. That's the reason we built twirl around generating PBR materials rather than complete concept images.
Keep the vehicle and change the material
If we're evaluating CMF, generating a completely new image of a car interior can actually make comparison more difficult. The geometry might change, the lighting might change, and the seats might suddenly have different stitching or proportions. Now we're comparing two different images rather than two material directions.
We'd rather keep the vehicle fixed. If we're considering a different textile, we want to generate the textile and put it onto our existing model. The same applies to a new polymer grain, leather alternative, or metal finish. That keeps the design controlled while letting the CMF change.
Generate the material itself
This is the workflow twirl is built around. Instead of asking a general image generator for a futuristic luxury car interior, we can describe the actual surface we want and generate that surface as a reusable asset.
twirl generates the description as a seamless PBR material rather than generating a new image of a car. We can apply it to the existing seat, render it under the same lighting, and compare it directly against the current textile. If we don't like it, we can change the prompt. Maybe the weave needs to be larger, the fibers warmer, or the surface more technical.
We're using generation to make CMF options, not to redesign the vehicle around us.
This works especially well early in the process
Early CMF exploration is usually full of uncertainty. We might know that the interior needs to feel warmer and more natural without knowing whether that means a textile, recycled composite, fine-grain polymer, or another material entirely. Building each of those possibilities manually can slow down the part of the process where we should be exploring broadly.
This is where twirl is most useful. We can generate several surfaces, put them onto the same geometry, and quickly eliminate the directions that don't work. Adobe, ArtCenter, and Rivian explored a digital-first material process in a transportation-design workshop where students created and compared multiple CMF directions using Substance 3D. Adobe's recap of the workshop is a good example of how digital material variation can support the design process without replacing physical material development.
Generation gives us another way to create the material options that feed into that process.
AI shouldn't invent material facts
There is a clear boundary here. A generated material can show us a visual direction. It can't tell us that the hypothetical textile contains 70 percent recycled fibers, that a polymer will pass an abrasion requirement, or that a supplier can manufacture the exact grain we've generated. Those claims need real data.
Research published in 2026 based on interviews with automotive CMF professionals looked specifically at how designers want AI to participate in this process. The researchers found interest in AI assistance around material information and circularity, but also concerns around reliability, traceability, authorship, and maintaining designer control. The Design Society paper is useful because it separates the appeal of AI assistance from the need for trustworthy material data.
That's a useful boundary for us too. twirl generates the visual material. We're not pretending the generated surface comes with manufacturing or sustainability facts that don't exist.
Build a library from the directions that work
Generation becomes more useful when the results don't disappear after the experiment. If we generate a textile that works well, it can become part of our twirl material library. The same material can come back for another concept, or we can use it as the starting point for a more developed physical direction.
Over time, this gives us a collection of surfaces shaped around the kind of work we actually do instead of a generic folder of textures. We can still browse the PBR material library when something already exists. When it doesn't, we can generate the material we actually have in mind.
AI can make the beginning of CMF much faster
We don't think AI needs to run the CMF process. A much more useful workflow is to give it one specific job. We have a surface idea, turn it into a material, put it onto the actual design, and decide whether it's worth pursuing.
If it isn't, we generate another direction. If it is, we can start doing the more expensive work: finding or developing a physical material, talking to suppliers, scanning it, measuring it, and validating it.
That's where twirl fits. It makes the funnel at the beginning much wider, so we can explore more material ideas before deciding which ones deserve the work that comes next.