Research map / 3D vision / Face and shape
TexVocab: Texture Vocabulary-conditioned Human 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 TexVocab, a texture-vocabulary avatar representation that back-projects multi-view video frames to SMPL UV texture maps, pairs them with body-part-wise pose keys, and conditions a NeRF on interpolated texture features for animatable human avatars.
Problem
Animatable human avatar modeling from multi-view RGB videos struggles to map driving poses to fine-grained pose-dependent appearance, since pose vectors carry no appearance information and latent-code or feature-line conditioning is blurry.
Approach
Back-project all training images onto the posed SMPL surface to build UV texture maps, store them as pose-keyed vocabulary entries, and encode poses per kinematic body part (rather than per joint) so that nearby-pose texture features are interpolated to condition a NeRF.
Key results
- On THUman4.0 subject00 novel poses, TexVocab reaches 32.09 PSNR, 0.9832 SSIM, and 18.79 FID versus 30.13 PSNR and 28.10 FID for PoseVocab.
- On DeepCap S2 against NeuralActor, PSNR improves from 23.531 to 26.325 and LPIPS drops from 0.066 to 0.017.
- The body-part-wise encoding beats global-pose and per-joint encodings on novel poses (PSNR 30.95 vs 29.59 and 29.23).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TexVocab instead of PoseVocab | TexVocab · THUman4.0 subject00, novel poses | PSNR: 30.13 -> 32.09 | Helped |
| use TexVocab instead of NeuralActor | TexVocab · DeepCap S2 | PSNR: 23.531 -> 26.325 | Helped |
| use body-part-wise embedding instead of global-pose or joint-structured embedding (ablation) | TexVocab · THUman4.0, 3 sequences, novel poses | PSNR: 29.59 -> 30.95; 29.23 -> 30.95 | Helped |
| use body-part-wise embedding instead of global-pose or joint-structured embedding (ablation) | TexVocab · THUman4.0, 3 sequences, novel poses | LPIPS*: 22.70 -> 19.06; 23.54 -> 19.06 | Helped |
| use texture maps from multi-view images instead of generator-predicted maps from SMPL normal maps (ablation) | TexVocab · THUman4.0, novel poses | visual quality | Helped |
Limitations
- Relies on SMPL inverse skinning, so loose clothing such as long dresses is not handled.
- Texture map gathering requires dense views; monocular or sparse-view inputs are not supported.
Experimental setup
Datasets: THUman4.0 ZJU-MoCap DeepCap AIST++
Benchmarks: THUman4.0 DeepCap
Metrics: PSNR SSIM LPIPS FID
Baselines: TAVA ARAH AniNeRF PoseVocab NeuralActor
Builds on: PoseVocab NeRF VolSDF SCANimate PixelNeRF
Closest papers by meaning
- Vid2Avatar-Pro: Authentic Avatar from Videos in the Wild via Universal Prior (CVPR 2025 · similarity 0.85)
- RAM-Avatar: Real-time Photo-Realistic Avatar from Monocular Videos with Full-body Control (CVPR 2024 · similarity 0.85)
- VINECS: Video-based Neural Character Skinning (CVPR 2024 · similarity 0.85)
- MPMAvatar: Learning 3D Gaussian Avatars with Accurate and Robust Physics-Based Dynamics (NeurIPS 2025 · similarity 0.85)
- Pose Modulated Avatars from Video (ICLR 2024 · similarity 0.85)
- HumanVid: Demystifying Training Data for Camera-controllable Human Image Animation (NeurIPS 2024 · similarity 0.85)
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.