OmniFaceRig: Fully Automatic Inner-Mouth-Aware Face Rigging Across Diverse 3D Character Topologies

How to read this

Category

fully automatic facial auto-rigging pipeline

Contributions

• End-to-end pipeline that turns a static, surface-only mesh with no oral cavity into a FACS rig of up to 155 blendshapes plus procedurally fitted teeth, gums, and tongue, with no manual landmarks or per-character templates
• Generalizes across humans, humanoids, and both long-muzzled (dogs, wolves, foxes) and short-muzzled (cats, bears, rabbits, tigers) animal topologies using topology-specific template selection
• Combines hybrid VLM plus CV riggability checking, multi-model face parsing and segmentation, dense keypoint template registration, and collision-aware blendshape transfer to reduce teeth-to-face intersection
• Releases Omni-Bench, a public benchmark of 1,000 biped 3D characters with FACS blendshapes and inner-mouth geometry across humans and animals, and reports 20 to 30 second end-to-end processing per asset on a single A100

Context

The paper positions itself against classical deformation transfer and parametric face models, which it says assume compatible topology, require manual landmarks, and target humans only, and against recent 3D asset generators such as Meta 3D AssetGen and its AssetGen2 follow-on that produce plausible-looking but non-rigged, non-animation-ready meshes. OmniFaceRig is meant to sit right after those generators in a pipeline, converting their static output into something an animator can pose. This entry lists no builds_on ancestors in the archive, but its stated lineage crosses auto-rigging methods with generative asset pipelines.

Correctness

The claims rest on the authors' own Omni-Bench benchmark, built and released alongside the method, plus quantitative results on rigging success rate, face detection recall, and inner-mouth penetration reported in sections not read in this pass. Because the benchmark is self-authored, performance on truly out-of-distribution creature topologies beyond the listed animal categories is unproven from the abstract alone, and the pipeline is explicitly scoped to the topology classes it defines templates for.

Clarity

The abstract and introduction are unusually concrete for a rigging paper, naming exact numbers (155 blendshapes, 1,000 characters, 20 to 30 seconds per asset, one A100) rather than vague claims, which lets a technical artist judge feasibility at a glance. The pipeline stages are named clearly, riggability checking, face parsing, template registration, blendshape transfer, even before reading the method section in depth.

How to read it

First pass, read the abstract and the contribution list to get the scope: which topologies, how many blendshapes, how fast. Second pass, read the Omni-Bench section and the pipeline overview to understand what riggability assessment screens out and what a template consists of for a non-human face. Third pass, for a production evaluation, read the ablations, latency breakdown, and limitations in the conclusion to see where the automatic pipeline still needs a human pass.