We’re ML researchers, but we started out making games. So when we look at today’s 3D generation models, we see both the progress and what’s still missing.
Flow-matching 3D models, texture synthesis, and automatic retopology have come a long way. But open the results in Blender and the problems show up fast: textures look like plastic and lack microdetail, retopology and rigging are inconsistent and far from robust, and every visual detail is baked into the mesh geometry instead of living on separate layers, which fuses the whole character into a single blob.
This makes the results difficult to edit, almost impossible to animate, and definitely not reusable. The current models produce output that might look convincing in a preview, but it doesn’t clear the bar for use in a real game.
A 3D artist works differently. Before modelling anything, they have a visual concept and break it down into meaningful parts: the body, hair, clothing, armour, and accessories. They understand that a cape is made of wool, a cuirass of bronze, and that each requires different geometry, materials, and topology. That understanding guides every step of the process.
At Anymate, we’re emulating this process. We call it Subject‑Aware Generation: a neuro-symbolic approach built on two sets of specialised neural networks, combining understanding with generation. On one side, vision models coordinated by a planning layer break the concept down into parts, positions, and materials. On the other, generative models turn that breakdown into geometry and textures, part by part. Textures follow the material, topology follows the anatomy, and every detail ends up as geometry, a normal map, or cards, whichever fits it best. The finished parts are assembled into one character instead of being carved out of a single fused mesh.
Explore the AnySeg demo {{ demoWhere }} to see how it all comes together {{ demoArrow }}
