Research map / 3D vision / Face and shape
FreeCloth: Free-form Generation Enhances Challenging Clothed Human Modeling
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 FreeCloth, a hybrid point-based framework that splits the body into unclothed, LBS-deformed, and free-form generated regions to model loose clothing like skirts and dresses.
Problem
Learning-based clothed human avatars rely on LBS of SMPL-like bodies, which fails for loose clothing far from the body (e.g., long dresses), producing split and fragmented artifacts.
Approach
Segment the body via a garment-specific clothing-cut map (SAM on normal maps) and model near-body clothing with LBS-based local deformation while generating distant loose clothing with an LBS-free, part-aware point cloud generator (modified SpareNet) conditioned on a part-based pose code.
Key results
- Overall FID 37.75 vs 39.02 for FITE and 57.87 for POP on ReSynth, the lowest among compared methods.
- On the challenging subject janett-025, FID 27.95 vs 35.81 for FITE (best).
- Human study: 63.4% of preferences for FreeCloth vs 29.9% FITE, 5.9% SkiRT, 0.8% POP; GPT-4o gives 56.0%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FreeCloth's hybrid modeling instead of FITE | Clothed human modeling · ReSynth, All | FID: 39.02 -> 37.75 | Helped |
| Use FreeCloth's hybrid modeling instead of FITE | Clothed human modeling · ReSynth, All | MSE (×10⁻²): 2.70 -> 2.61 | Helped |
| Use only LBS-based deformation instead of the hybrid model (ablation) | FreeCloth · ReSynth, loose clothing | visual quality | Hurt |
| Use only free-form generation for the lower body instead of the hybrid model (ablation) | FreeCloth · ReSynth, articulated lower body | visual quality | Hurt |
| Remove collision loss (ablation) | FreeCloth · ReSynth, loose clothing | body-clothing intersections | Hurt |
| Remove part-aware pose feature encoding (ablation) | FreeCloth generator · ReSynth, posed skirts | pose alignment | Hurt |
| Remove the clothing-cut map and use the default UV map (ablation) | FreeCloth · ReSynth, long dresses | visual quality | Hurt |
Surprising findings
- Entirely LBS-free generation of the full lower body yields noisy, discontinuous results, so the hybrid design is needed.
Limitations
- Evaluated only on the synthetic ReSynth dataset with five subjects in skirts and dresses.
- Relies on SAM-based segmentation of loose regions from rendered normal maps to define the clothing-cut map.
- Geometry only; texture rendering via 3DGS is left to future work.
Experimental setup
Datasets: ReSynth
Benchmarks: ReSynth
Metrics: FID MSE of rendered normal maps human preference rate GPT-4o preference
Baselines: POP SkiRT FITE
Builds on: POP CloSET SkiRT SpareNet SMPL-X
Closest papers by meaning
- DLCA-Recon: Dynamic Loose Clothing Avatar Reconstruction from Monocular Videos (AAAI 2024 · similarity 0.90)
- DAGSM: Disentangled Avatar Generation with GS-enhanced Mesh (CVPR 2025 · similarity 0.89)
- Design2Cloth: 3D Cloth Generation from 2D Masks (CVPR 2024 · similarity 0.89)
- Garment Recovery with Shape and Deformation Priors (CVPR 2024 · similarity 0.88)
- DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images (NeurIPS 2026 · similarity 0.88)
- GarFast: Realistic and Fast Garment Transfer with a Simplified Parser-Free Approach (AAAI 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.