AI character animation generator, from rig to looping clips

Meshy's rigging documentation lists what auto-rigging suits. Standard humanoid, bipedal characters with clearly defined limbs and body structure, textured, with the face pointing toward +Z, the glTF forward direction. Untextured meshes, non-humanoid creatures and humanoids with unclear limb structure are listed as unsuitable, and Meshy's API rejects models over 300,000 faces when rigging from one of its own generation tasks, so keep character meshes well under that. Meshy takes the approximate height of the character in meters, 1.7 by default. Studio uses that default, and your agent can pass the real height through the MCP.

Two practical consequences. Generate characters in an A or T pose with hands away from the body, which the 3D tab's Character intent does for you. And keep quadrupeds, creatures and props with moving parts out of this pipeline, because both the rigger and the text-to-motion model are humanoid only.

Hunyuan Motion is Tencent's HY-Motion model served through fal. You describe an action and it returns a skeletal clip of 0.5 to 12 seconds as FBX, on its own humanoid skeleton with SMPL-H style bone names such as Pelvis, L_Hip and R_Shoulder. It does not know your character, so the clip is fitted to your rig when you export from the Animation tab.

Tencent's own guidance is the best prompt advice. Keep prompts under 60 words and in English, and describe the action rather than the character's looks, clothing, mood, the scene or the camera. One humanoid only, since multi-person interactions and animals are listed as unsupported. Short single actions with a clear order, such as a person running then taking a big leap, come back cleanest.

• Write the verbs in order, for example "steps forward, swings the right arm overhead, then recovers to a guard stance".
• Set the length to the action. A two to three second emote needs no twelve second clip.
• Generate a few takes and keep the best one; every run is a new clip.
• Plan to trim and loop by hand, because looping and in-place motion are not supported by the model.

Clips from the library already share your character's skeleton, so they need no retargeting. Summer merges them onto the base rig by node name and validates the result; if a clip's skeleton does not match, the job fails instead of producing a twisted animation. Mixing clips from two separate rig runs on one character is refused for the same reason.

For clips from elsewhere, Godot 4 retargets at import time. In the Advanced Import Settings of the scene, select the Skeleton3D, create a BoneMap with SkeletonProfileHumanoid, and let it match bone names. Godot's docs call Overwrite Axis the most important option for sharing animations between models, Fix Silhouette helps when one model is in an A pose and the other in a T pose, and Normalize Position Tracks stops stride slipping between characters of different heights. If automatic matching misses bones, assign them by hand in the BoneMap.

Summer's packages and exports are GLB, which is Godot's recommended format and carries skinning and animation in one file. Hunyuan Motion clips arrive as FBX, and Godot 4.3 and later import FBX directly through ufbx, so you can also bring them into a project without converting.

Loops need attention in any pipeline. Godot sets the loop flag on import only for clips whose names start or end with loop or cycle. Summer's character package records which clips loop in its manifest, and the default controller sets loop mode on idle, walk and run when the character starts. Tencent's HY-Motion documentation lists looping and in-place animation as unsupported, so a text-to-motion clip needs trimming to a clean cycle before you loop it.

There is no video-to-mocap path, no facial animation and no non-humanoid rigging. Text-to-motion is not retargeted automatically when it is generated; that happens at export in the Animation tab. For hand-keyed polish, the clips are standard GLB and FBX that open in Blender.