I used an image-generation AI to create clear three-view reference images without textures.
I used Tripo3D's multi-view generation to create a high-poly model directly, with one million polygons.
Once I had the high-poly model, I manually retopologized it to make a low-poly version and completed the remaining material workflow.
I found a problem: even with a very high polygon count, the AI-generated model still lacked detail. Because the mesh was triangulated, adding subdivisions could not resolve these shortcomings. They directly affected the quality of the baked normal map. At this stage, these AI-generated high-poly models cannot be used directly in production.
AI does improve efficiency, though. For example, it can help modelers quickly establish an object's overall shape and generate a quad mesh, which they can then sculpt by hand.