Facial Rig & Shape Keys

The execution model is in blender-core. The body rig is in blender-character-rigging.

Choose the approach

Procedure (blendshapes)

1. The head mesh needs clean loops around the eyes and mouth (blender-retopology-uv), with eyes, teeth and tongue as separate objects parented or weighted to bones.
2. scripts/shapekeys.py -- --mesh Head --create arkit (or --create visemes for the 15 Oculus visemes) creates empty, correctly named keys.
3. Sculpt each key (live: sculpt mode on the active key, use_shape_key_edit_mode). Symmetric expressions can be sculpted once, then --split eyeBlink gives eyeBlinkLeft/Right with a soft blend across X=0 (--falloff m).
4. --mirror browDownLeft builds browDownRight (topology-independent mirror).
5. --report lists unsculpted keys, the max offset per key, and missing ARKit names. A shape is done only when every key moves and nothing is left over.
6. Optional: --drivers FaceCtrl adds a controller with 0–1 custom properties driving each key, for keyframing in Blender.
7. Check the extremes and combinations (jawOpen + mouthSmile, eyeBlink + eyeSquint) with look. Fix them with corrective combination keys driven by the product of two keys.

Shape key rules

• Shapes are deltas from Basis. Never change the Basis topology after creating keys (adding or removing verts destroys every key).
• One action per key. No head or neck movement inside face keys (bones do that). Keep eyelids sliding over the eye surface, not through it (check with blender-asset-qa zfighting on lid/eye).
• Range 0–1. slider_max can be raised for exaggeration, but engines clamp to 0–1.
• Relative keys (use_relative=True, the default) are the only kind that exports.
• Name exactly per spec (case-sensitive): eyeBlinkLeft, not EyeBlink_L. Engines map by name.
• Keys on meshes with Subdivision: sculpt at level 0, and keep the subdivision modifier after the shape keys (they're evaluated first anyway).
• Applying modifiers on a mesh with shape keys fails. Use modifier_apply_as_shapekey for deform modifiers, or apply them before creating keys.

Correctives and combination shapes

• Corrective for a bone pose: sculpt in the posed state, then drive with a TRANSFORMS variable (ROT_X, LOCAL_SPACE, mapped to 0–1).
• Combination: key jawOpen_mouthSmile with driver SCRIPTED expression a*b, from the two key values (SINGLE_PROP key_blocks["jawOpen"].value). Simple expressions don't need Python auto-run.

Engine notes

• Unity: FBX blendshapes. In the Model importer, set Blend Shape Normals = Calculate (or Import if smoothing was exported). ARKit names map to the AR Foundation face-tracking samples.
• Unreal: morph targets ("Import Morph Targets"). Live Link Face / MetaHuman Animator use ARKit names, and the Pose Asset maps the curves. No leaf bones.
• Roblox dynamic heads: FACS-based poses with Roblox-specific names, authored as animation frames on a head rig (Creator Hub "Dynamic heads" spec). Build from the official Blender template. ARKit sculpts can be reused by renaming/mapping. Head ≤ 4,000 tris.
• VRM / web (three-vrm, ReadyPlayerMe-style): ARKit 52 + visemes as glTF morph targets (export_morph=True). Keep the morph-normal export on only if shading needs it (it doubles the size).
• Drop unused keys before export. Every key costs memory per vertex (glTF sparse accessors help).

Gotchas

• ob.shape_key_add(from_mix=False) copies the Basis. from_mix=True bakes the current mix (useful for combos, dangerous by accident).
• shape_key_mirror(use_topology=False) needs a mesh symmetric on X. Run it on the active key (the script handles this).
• Driver variables on shape keys use the Key ID (ob.data.shape_keys), not the mesh: fc = kb.driver_add("value").