Research map / 3D vision / Face and shape
DAGSM: Disentangled Avatar Generation with GS-enhanced Mesh
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.
Generates disentangled text-driven 3D avatars by sequentially creating a body and individual garments as GS-enhanced meshes (2DGS bound to mesh), with SAM-based filtering and view-consistent texture refinement.
Problem
Text-driven 3D avatar generation usually models the body and all garments as one holistic model, which prevents clothing replacement, limits user control, and yields unrealistic animation and low-quality textures.
Approach
Represent each body/garment part as a GS-enhanced mesh (2D Gaussians attached to mesh triangles), generate body first then garments sequentially with SAM-based Gaussian filtering for separation, and refine textures with cross-view attention and incident-angle-weighted denoising.
Key results
- CLIP ViT-L/14 similarity 28.8 for DAGSM vs 27.3 for HumanGaussian and 24.2 for DreamWaltz
- User study (50 participants): 74.0% votes for visual quality, 76.8% for text alignment, 68.4% for animation realism
- Body generation takes about 60 minutes on one NVIDIA A100 40GB; garment generation and texture steps take 30 and 20 minutes
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAGSM rather than the strongest baseline, HumanGaussian | 10 generated avatars · 10 generated avatars | CLIP similarity (ViT-L/14): 27.3 -> 28.8 | Helped |
| Use DAGSM rather than the strongest baseline, HumanGaussian | 10 generated avatars · 10 generated avatars | CLIP similarity (ViTbigG-14): 43.9 -> 45.8 | Helped |
| Use DAGSM rather than the strongest baseline, SO-SMPL | 10 generated avatars · User study, 10 generated avatars | visual quality votes: 16.6% -> 74.0% | Helped |
| Use DAGSM rather than the strongest baseline, SO-SMPL | 10 generated avatars · User study, 10 generated avatars | text alignment votes: 11.6% -> 76.8% | Helped |
| Use DAGSM rather than the strongest baseline, SO-SMPL | 10 generated avatars · User study, 10 generated avatars | animation realism votes: 20.4% -> 68.4% | Helped |
| Add SAM-based filtering to remove non-garment Gaussians (ablation) | DAGSM garment generation · Garment generation | body-garment separation | Helped |
| Add view-consistent refinement (ablation) | DAGSM texture generation · White lace dress texture | texture quality | Helped |
| Add cross-view attention during refinement (ablation) | DAGSM texture refinement · White lace dress texture | texture style consistency | Helped |
| Add incident-angle-weighted denoising (IAW-DE) (ablation) | DAGSM texture refinement · White lace dress texture | texture blurriness | Helped |
Limitations
- TSDF mesh extraction only reconstructs garment surfaces, omitting internal structures such as pockets
- SO-SMPL-style body-topology constraints are avoided only partially; evaluation limited to 10 avatars
Experimental setup
Datasets: AIST++
Benchmarks: 10 generated avatars with text prompts (in-house)
Metrics: CLIP similarity (ViT-L/14, ViTbigG-14) user study preference rate
Baselines: DreamWaltz HumanGaussian SO-SMPL
Builds on: DreamFusion 2D Gaussian Splatting SMPL-X RFDS loss TIP-Editor GaMeS
Closest papers by meaning
- FreeCloth: Free-form Generation Enhances Challenging Clothed Human Modeling (CVPR 2025 · similarity 0.89)
- CtrlAvatar: Controllable Avatars Generation via Disentangled Invertible Networks (AAAI 2025 · similarity 0.89)
- Design2Cloth: 3D Cloth Generation from 2D Masks (CVPR 2024 · similarity 0.88)
- SimAvatar: Simulation-Ready Avatars with Layered Hair and Clothing (CVPR 2025 · similarity 0.87)
- DreamFit: Garment-Centric Human Generation via a Lightweight Anything-Dressing Encoder (AAAI 2025 · similarity 0.86)
- En3D: An Enhanced Generative Model for Sculpting 3D Humans from 2D Synthetic Data (CVPR 2024 · similarity 0.86)
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.