Research map / 3D vision / Face and shape
LUCAS: Layered Universal Codec Avatars
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LUCAS, a layered universal prior model for codec avatars that separates hairless head and hair into parallel mesh branches, enabling real-time mesh avatars and better Gaussian anchor geometry.
Problem
Photorealistic 3D head avatars struggle to model dynamic face-hair interactions and generalize across identities; prior universal prior models treat hair as part of a single mesh, causing misaligned geometry and artifacts during expressions and head motion.
Approach
Extends a universal Pixel Codec Avatar (uPiCA) to a layered design with two parallel hypernetworks, expression encoder shared across branches, and separate face and hair decoders with multi-mesh rendering; hair decoders also take head pose, neck pose and expression code. Dehairing via EM-based linear deformable bald head model provides bald geometry supervision.
Key results
- LUCAS (gs) reaches 34.56 PSNR / 0.9201 SSIM / 0.2394 LPIPS on cross-subject setting vs 33.12 PSNR for URAvatar (gs).
- LUCAS (mesh) improves cross-subject PSNR from 32.56 (uPiCA) to 33.03 and per-subject PSNR from 32.05 (PiCA) to 33.52.
- Real-time mesh avatar runs at 45 FPS on mobile according to Figure 1 caption.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LUCAS (mesh) instead of PiCA (mesh) | Per-subject avatar reconstruction · Per-subject optimization | PSNR: 32.0512 -> 33.5211 | Helped |
| use LUCAS (gs) instead of URAvatar (gs) | Universal Gaussian avatar reconstruction · Cross-subject optimization | PSNR: 33.1227 -> 34.5579 | Helped |
| add the expression code for hair (ablation) | LUCAS · Training subjects | PSNR: 34.1014 -> 34.4981 | Helped |
| add the expression code for hair (ablation) | LUCAS · Unseen subjects | PSNR: 31.9128 -> 32.5847 | Helped |
| add hair segmentation regularization (ablation) | LUCAS · Training subjects | PSNR: 34.0285 -> 34.4981 | Helped |
| add hair segmentation regularization (ablation) | LUCAS · Unseen subjects | PSNR: 31.7964 -> 32.5847 | Helped |
Limitations
- Struggles with extreme hair deformations.
- Unseen poses during zero-shot driving can degrade long hair deformation.
- Trained on a limited set of hairstyles; relighting not addressed.
Experimental setup
Datasets: in-house multi-view capture of 76 identities (110 cameras)
Benchmarks: in-house multi-view capture held-out subjects
Metrics: PSNR SSIM LPIPS
Baselines: PiCA (mesh) uPiCA (mesh) URAvatar (gs)
Builds on: Pixel Codec Avatars URAvatar 3D Gaussian Splatting Universal Prior Model (Cao et al.)
Closest papers by meaning
- MeGA: Hybrid Mesh-Gaussian Head Avatar for High-Fidelity Rendering and Head Editing (CVPR 2025 · similarity 0.91)
- HERA: Hybrid Explicit Representation for Ultra-Realistic Head Avatars (CVPR 2025 · similarity 0.88)
- Relightable Gaussian Codec Avatars (CVPR 2024 · similarity 0.88)
- DGH: Dynamic Gaussian Hair (NeurIPS 2025 · similarity 0.88)
- Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians (CVPR 2024 · similarity 0.88)
- Synthetic Prior for Few-Shot Drivable Head Avatar Inversion (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.