RegHead: Non-Humanoid Head Blendshapes via Feed-Forward Registration

2026-07-13 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

RegHead is a novel framework designed to construct semantic blendshape sets for animatable non-humanoid head avatars, addressing the high cost and complexity of current methods. Traditional blendshape creation is hindered by scarce expression-consistent supervision, lack of correspondence in generated 4D assets, and highly localized facial motion. RegHead overcomes these challenges by introducing three key components: a large-scale dataset of non-humanoid identities with a shared expression vocabulary, expanded through fine-tuned image editing; a dense stochastic anchor motion representation specifically for localized facial deformations; and a fast feed-forward registration model. This model converts unregistered expression meshes into a corresponded blendshape basis by predicting anchor-based deformations from a neutral shape. Experiments demonstrate that RegHead produces higher-fidelity expression meshes and operates significantly faster than optimization-based baselines, enabling real-time retargeting from human face tracking signals to non-humanoid characters.

Key takeaway

For 3D animators or game developers building non-humanoid characters, RegHead offers a significantly faster and higher-fidelity solution for generating semantic blendshapes. If you are struggling with the cost and complexity of traditional rigging for diverse non-humanoid avatars, consider adopting this feed-forward registration approach. It allows you to achieve real-time retargeting of human facial tracking data onto your custom characters, streamlining animation workflows and expanding creative possibilities for unique digital identities.

Key insights

RegHead enables efficient, high-fidelity semantic blendshape creation for non-humanoid avatars using a novel dataset, anchor representation, and feed-forward registration.

Principles

• Semantic blendshapes offer low-dimensional, interpretable animation.
• Localized facial motion requires specialized representation.
• Feed-forward registration can outperform optimization for speed.

Method

RegHead constructs blendshapes by expanding an artist-rigged library via image editing for a dataset, then uses a dense stochastic anchor motion representation with a feed-forward registration model to predict deformations from neutral shapes.

In practice

• Real-time retargeting from human face tracking.
• Animating non-humanoid characters with human expressions.
• Creating diverse non-humanoid avatar libraries.

Topics

• RegHead
• Non-Humanoid Avatars
• Semantic Blendshapes
• Facial Animation
• 3D Registration
• Real-time Retargeting

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• RigPAPR: Rig-Based Animation of Static Neural Point Clouds from a Fixed-Viewpoint Video — PAPR's pixel recomposition avoids joint artifacts in rig-based neural point cloud animation.
• Ex-Omni: Enabling 3D Facial Animation Generation for Omni-modal Large Language Models — Ex-Omni integrates 3D facial animation into OLLMs by decoupling semantic reasoning from temporal generation using speech units and TQGF.
• Adaptive Identity Anchoring: Closed-Loop Keyframe Placement for Synthetic Paired Supervision in Video Face Swapping — Video face swapping can overcome identity drift and texture loss via adaptive anchoring and reality-referenced texture restoration.
• 👩‍🦰Deformable 3D Hair👩‍🦰 👉Xi’an Jiaotong University unveils a novel method... — A new method reconstructs decoupled 3D Gaussian head avatars from a single image, enabling dynamic hairstyle transfer.
• MeshLoom: Feed-Forward Non-Rigid Registration of Mesh Sequences — MeshLoom uses a feed-forward network with topology-aware encoding and multi-modal fusion for efficient non-rigid mesh sequence registration and interpolation.
• Pro-Pose: Unpaired Full-Body Portrait Synthesis via Canonical UV Maps — Pro-Pose generates high-fidelity, reposed human avatars from single images by decoupling pose from texture in UV space.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.