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

Abstract

Overview

We present RegHead, a framework for constructing semantic blendshape sets for animatable non-humanoid head avatars. With a fixed expression vocabulary, semantic blendshapes provide a low-dimensional and interpretable animation interface and support cross-identity retargeting. Building such blendshape sets remains expensive because (i) expression-consistent supervision is scarce, (ii) generated 4D assets typically lack correspondence, and (iii) facial motion is highly localized. We propose (1) a large-scale dataset of non-humanoid identities paired with a shared expression vocabulary, obtained by expanding a small artist-rigged library via fine-tuned image editing; (2) a dense stochastic anchor motion representation tailored to localized facial deformations; and (3) a fast feed-forward registration model that converts unregistered expression meshes into a corresponded blendshape basis by predicting anchor-based deformations from the neutral shape. Experiments show that our approach produces higher-fidelity expression meshes than baselines, while running orders of magnitude faster than optimization. We further demonstrate real-time retargeting from human face tracking signals to non-humanoid characters, capturing both head pose and localized facial motions.

The Dataset

An expression-aligned population of non-humanoid heads

Each identity is one distinct animal head rendered across 10 fixed facial expressions. Because every identity shares the same expression vocabulary, the dataset provides aligned cross-expression pairs for a fixed identity — exactly the supervision needed for blendshape learning, expression editing, and identity-preserving animation.

33,724

Unique identities

10

Expressions per identity (100% complete)

337,240

Total images

2

Resolutions · 512 & 1024 px

The 10 expressions, one identity

Same expression vocabulary for all 33,724 identities

Composition & diversity

Identities are sampled as combinations of an animal (~270 species), a render style (4), a facial feature (~38), and 1–2 accessories (~57) — producing broad visual diversity across the population. Every image is model-generated; no real photographs are included.

~59% of identities are newly generated for this release; the remaining ~41% come from an earlier batch using the same expression vocabulary and models.

RENDER-STYLE DISTRIBUTION

Video Game NPC25.5%

Plastic Toy Render25.0%

Photorealistic25.0%

Computer Animated Film-Style 3D24.5%

Styles are sampled near-uniformly across the four render looks.

How it was generated

A three-stage generative pipeline: create a base head, neutralize it into a clean reference, then edit it into each expression with chained, expression-specific adapters.

1

Text → image

A seeded prompt (animal × style × feature × accessories) renders the base identity head.

OpenAI gpt-image-1

2

Neutralize

The base head becomes a clean floating-head reference on a white background — the open_m_open_e frame.

OpenAI image model

3

Expression edits

LoRA adapters chain-edit the reference into the remaining 9 expressions, keeping identity fixed.

Qwen-Image + LoRA

Two releases — same content, two resolutions

	512 release	1024 release
Resolution	512 × 512	1024 × 1024
Image format	JPEG q95	JPEG q95
Identities	33,724	33,724
Images	337,240	337,240
Shards (.tar)	11	34
Total size	18.96 GiB	58.79 GiB
Best for	fast prototyping, light training	high-res training & evaluation

Identical identity sets, expressions, metadata, and per-sample layout — only the resolution and size differ. Train on 512, scale to 1024 without changing any data-loading code.

Load in one shard, one sample = one identity

# packaged as WebDataset .tar shards — pip install webdataset pillow import glob, json, webdataset as wds urls = sorted(glob.glob("reghead_download/data/*.tar")) EXPR = ["open_m_open_e", "halfo_m_o_e", "close_m_o_e", "close_m_halfo_e", "close_m_0_25o_e", "close_m_0_75o_e", "close_m_close_e", "close_m_smile", "raise_eyebrows", "frown"] def parse(s): meta = json.loads(s["json"]) if isinstance(s["json"], (bytes, bytearray)) else s["json"] imgs = {e: s[e + ".jpg"] for e in EXPR if e + ".jpg" in s} return {"identity": meta["identity"], "images": imgs} ds = wds.WebDataset(urls).decode("pil").map(parse) # imgs are PIL.Image

Method

Feed-forward registration

RegHead overview: from 2D observations to blendshapes to real-time retargeting — **Figure 1.** RegHead converts semantically labeled expression observations into a corresponded semantic blendshape set for non-humanoid heads in a single feed-forward pass. The resulting blendshapes support real-time animation and retargeting via a fixed expression vocabulary.

Results

Qualitative results

All results below are produced by RegHead. Each clip pairs the input with our output (baselines cropped out). Videos autoplay, muted, when scrolled into view.

Neutral → target interpolation

Linear interpolation in the registered blendshape basis, morphing the neutral shape toward each target expression.

Neutral → Closed mouth, open eyes. Each identity is shown as an input / ours pair, left to right: African Lion, Bald Eagle, Chihuahua Dog.

Neutral → Closed mouth, closed eyes. Each identity is shown as an input / ours pair, left to right: Alligator, Barracuda, Beaver.

Neutral → Smile. Each identity is shown as an input / ours pair, left to right: Lynx, Mongoose, Parrot.

Neutral → Frown. Each identity is shown as an input / ours pair, left to right: Mouse, Orangutan, Pelican.

Retargeting

Real-time retargeting from a human face-tracking signal to non-humanoid characters, transferring head pose and localized facial motion.

Input 1. Leftmost is the driving human performance; the remaining three are our real-time retargeting to Maine Coon, Mole, Prairie DogCockatoo.

Input 2. Leftmost is the driving human performance; the remaining three are our real-time retargeting to Maltese Dog, Mouse, Musk Ox.

Input 3. Leftmost is the driving human performance; the remaining three are our real-time retargeting to Lemur, Mouse, Poodle.

Citation

BibTeX

@inproceedings{reghead2026, title = {RegHead: Non-Humanoid Head Blendshapes via Feed-Forward Registration}, author = {Author, One and Author, Two and Author, Three}, booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)}, year = {2026} }