Research map / 3D vision / Face and shape
AttriHuman-3D: Editable 3D Human Avatar Generation with Attribute Decomposition and Indexing
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 AttriHuman-3D, an editable 3D human avatar GAN that generates all clothing/body attributes in a shared space-attribute field decomposed into six feature planes, with implicit attribute indexing and orthogonality regularization for part-level editing.
Problem
Editable 3D-aware human avatar generation: existing editable 3D GANs either lack high-accuracy local attribute editing or require many parallel generators, which is computationally expensive; datasets also show style entanglement between attributes.
Approach
Formulate generation as a 4D space-attribute field decomposed (HexPlane-style) into six feature planes generated by one StyleGAN2-based 2D CNN, with a learned implicit attribute index (MLP on one-hot labels, L2-normalized) and an orthogonal projection regularization so attribute features can be extracted and swapped independently; plus hyper-latent (gender) conditioning and attribute-specific boun
Key results
- Full model reaches FID 16.85 and FID_edit 17.43 on DeepFashion, versus 15.91 for EVA3D (no editing) and 27.69 / 30.31 for CNeRF.
- Ours uses 9G memory and 59M parameters at 0.09 s, compared with 16G and 0.23 s for CNeRF; the 11-generator base model runs out of memory.
- Removing orthogonal projection regularization raises FID from 16.85 to 19.81 and fixed indexing raises it to 21.20.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace fixed indexing with implicit attribute indexing (ablation) | AttriHuman-3D · DeepFashion | FID: 21.20 -> 16.85 | Helped |
| add orthogonal projection regularization (ablation) | AttriHuman-3D · DeepFashion | FID: 19.81 -> 16.85 | Helped |
| add hyper-latent training strategy (ablation) | AttriHuman-3D · DeepFashion | FID: 17.24 -> 16.85 | Helped |
| add attribute-specific sampling (ablation) | AttriHuman-3D · DeepFashion | FID: 17.02 -> 16.85 | Helped |
| add hyper-latent training strategy (ablation) | AttriHuman-3D · DeepFashion | PCK: 89.63 -> 89.91 | Hurt |
| add attribute-specific sampling (ablation) | AttriHuman-3D · DeepFashion | PCK: 89.97 -> 89.91 | Hurt |
Limitations
- Experiments are limited mainly to the DeepFashion dataset (11,903 images).
- Generation quality is slightly below EVA3D on FID (16.85 vs 15.91), and editing FID rises a little.
- Style entanglement between attributes is only addressed heuristically through hyper-latent and bounding-box strategies.
Experimental setup
Datasets: DeepFashion
Benchmarks: DeepFashion
Metrics: FID FID_edit PCK Pseudo Depth memory parameters time per iteration
Baselines: EG3D StyleSDF CNeRF EVA3D
Builds on: EG3D IDE-3D HexPlane TensoRF EVA3D StyleGAN2
Closest papers by meaning
- PERSE: Personalized 3D Generative Avatars from A Single Portrait (CVPR 2025 · similarity 0.88)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.87)
- 3DHumanEdit: Multi-modal Body Part-aware Conditioning Information Integration for 3D Human Manipulation (AAAI 2025 · similarity 0.87)
- PEGASUS: Personalized Generative 3D Avatars with Composable Attributes (CVPR 2024 · similarity 0.86)
- GeneAvatar: Generic Expression-Aware Volumetric Head Avatar Editing from a Single Image (CVPR 2024 · similarity 0.85)
- Reference-Based 3D-Aware Image Editing with Triplanes (CVPR 2025 · 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.