Skip to content

Rigged GLB export: 0.01 object scale with 100x baked data, bundled Icosphere helper, and foot skin-weight stretching #3

Description

@wangcan26

Summary

When downloading a rigged model as GLB via meshy_download_model (task_type='rigging') and importing it into Blender, the exported file has a broken transform/scale setup and a few extra artifacts. The rendered result is fine, but the scene is very hard to work with and easily looks "empty/broken" to users (and to AI agents driving Blender via MCP). These all originate in the Meshy export, which this MCP server is the front door for.

Tested pipeline: meshy_image_to_3d (pose_mode='t-pose') → meshy_rig (free walking/running) → meshy_download_model (task_type='rigging', GLB) → Blender import_scene.gltf.

Issues

1. Extreme object scale (0.01) with 100× baked data (main issue)

In the rigged GLB, both the mesh object (char1) and the Armature have object scale = 0.01, while the mesh vertices / bone data are baked 100× larger.

  • Object_scale × data → correct world size (character ≈ 1.6–1.7 m), so F12 render is correct.
  • But because the bone data is 100×, the Armature's viewport display is ~10 × 12 × 11 m, completely enclosing/occluding the 1.6 m character.
  • Net effect: open the imported file and the viewport shows a giant octahedron/"sphere" (the bone display) and the character appears tiny or hidden — users think the import failed or the scene is empty.

Expected: object scale should be 1.0 (or a clean uniform value) with un-baked data, so the armature display and the mesh share a sane real-world scale out of the box.

2. Bundled hidden helper object / collection

The GLB ships with a hidden Icosphere (~2 m, hide_render=True, hidden in outliner) and import creates a glTF_not_exported collection to hold such helpers. This is unexpected clutter in a "character" export.

3. Auto-rig foot skin weights cause stretching

With the free auto-rig walking/running animation, the feet/shoes stretch into long smears on lift/plant frames (visible from the side, not the front), with minor foot-through-floor penetration (foot world z down to ≈ -0.02). Foot vertices appear to be pulled by the lower-leg bones (LeftLeg/RightLeg); the end-effector (foot/toe) auto-weights are inaccurate. Re-mesh + re-rig helps but extreme stride frames still stretch.

4. (Minor / possibly by design) No root motion

The bundled walking/running animations have no root motion — the character walks in place. This is convenient for framing but worth documenting; please clarify whether root-motion variants are possible.

Environment

  • meshy-mcp-server (latest), Blender 4.x/5.x, glTF import.
  • Models: humanoid character, image-to-3d → rig.

Ask

Primarily 1 (the 0.01-scale / 100×-baked transform on rigged GLB exports) and 3 (foot auto-weights) are the impactful ones. Could the rigged GLB be exported with a unit object scale and clean (applied) transforms, drop the hidden Icosphere helper, and improve end-effector weights? Happy to provide a sample task ID / GLB if useful.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions