New Organic Mode: AI game assets, auto rigging and animation

We deployed a new pipeline behind Organic Mode, the part of GrandpaCAD that makes characters, creatures, busts and anything else you'd describe by how it looks. I'm not going into how it works. The short version: meshes come out a lot more detailed, and it's far more flexible than anything we had before.

Just say what you want

There are no settings to learn. You describe the model, and the AI works out the polygon count, the textures, the size and the file format from what you wrote.

Make me a golem character in a T pose with 8000 polygons

That's (almost) the whole prompt. I added "then rig it" at the end. You get a game-ready GLB with exactly 8,000 triangles, a color texture, a normal map and a 35-bone skeleton.

A few more that work:

• "A low-poly treasure chest for a mobile game, under 2000 triangles"
• "A highly detailed bust of a bearded dwarf for a Blender render"
• "A cartoon frog figurine, 8 cm tall, in 4 colors for my AMS"
• "A stylized oak tree prop, 5000 polygons, 2K textures"
• "Same golem, but 20,000 polygons this time" (it reuses the shape it already made, so this one's quicker)

Textures included

Game and render models come with an albedo (color) map and a normal map baked in. That's enough for most games and quick renders.

Want full PBR, with roughness, metallic and height maps too? Generate the material with PATINA on fal and apply it in Blender, Unity or Unreal.

AI auto rigging and animation (alpha)

This part is alpha. It works often enough that I'd rather you try it than wait, but expect some misses.

Auto rigging

Just ask: "rig it". GrandpaCAD puts a skeleton inside the model and skins the mesh to it (every vertex gets weights for the bones that move it), and you get the same GLB back, textures untouched, ready to pose in Blender or drop into Unity, Unreal or Godot. Humans, animals, creatures and robots all work. It takes about 20 to 40 seconds and costs one generation, same as each animation request.

Auto animation

Once a character has a rig, ask for motion in plain words: "make it dance", "a walk cycle", "wave hello". Each request adds a named animation clip to the GLB, and the viewer plays it right away.

One prompt, "Make me a stylized cartoon adventurer character in a T pose as a textured game asset GLB, about 10000 polygons. Then rig it, and then animate it with a dance clip", gave me this 7 second loop:

What to expect from the alpha:

• Bones get generic names (bone_0, bone_1...), not LeftArm or Spine. Retargeting to a Mixamo or Unreal skeleton needs a manual mapping.
• No finger or face rigs, and no finger control in the motion.
• Generate the character standing in a T pose or A pose, limbs apart from the body. A sitting or crouching figure rigs and animates badly, and limbs fused to the torso share one bone.
• Walking, running, jumping, turning, idling, waving and dancing work best. Pose-type prompts (sit, lie down, roll over) often come out contorted.
• Clips are built from 2 second chunks, up to 10 seconds each, and nothing blends between clips (your game engine does that).
• Rigs and animations live in GLB only. Ask for a 3MF or STL and they're dropped.

If you're curious what does the rigging: it's TokenRig from VAST AI Research ("SkinTokens: A Learned Compact Representation for Unified Autoregressive Rigging"), MIT licensed, running on our own GPU deployment. It predicts the skeleton and the skin weights in one go from the mesh alone, which is why it works on creatures as well as humans.

Where to try it

It's live in the new chat, part of the GrandpaCAD 2.0 beta. GrandpaCAD for Games takes a dragon from prompt to rig, and animates a golem from a single message. If you want the longer version (formats, print-specific defaults, editing your own files), it's in the games, renders and printing post.