Research map / 3D vision / Face and shape
Design2Cloth: 3D Cloth Generation from 2D Masks
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.
Design2Cloth trains a triplane-based implicit UDF garment GAN on a new 2k-scan real-world cloth dataset (DigitalMe) so users can generate and reconstruct 3D garments from 2D masks.
Problem
3D garment generation is limited by garment-specific models or synthetic training data that lack wrinkles and realism, and existing conditioning (UV maps, point clouds) is impractical for users.
Approach
A GAN with a triplane implicit UDF generator conditioned on a 2D visibility mask and SMPL shape, using a dual-resolution point-cloud discriminator (uniform branch plus high-curvature branch) to enforce wrinkles, trained on a large real-world scan dataset.
Key results
- On DigitalMe, Design2Cloth reaches CD 0.12e-2 and NC 0.98 versus DrapeNet's 0.56e-2 and 0.96.
- In a 50-person user study, generated garments scored 7.2 realism versus 4.6 for DrapeNet and 7 for Cloth3D ground truth.
- On CustomHumans single-image reconstruction, CD is 0.12e-2 for the proposed method versus 0.40e-2 for SMPLicit and 1.14e-2 for DrapeNet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Design2Cloth instead of DrapeNet for cloth reconstruction | Cloth generation on Cloth3D · Cloth3D | CD (×10−4): 0.36 -> 0.18 | Helped |
| Use Design2Cloth instead of DrapeNet for cloth reconstruction | Cloth generation on Cloth3D · Cloth3D | NC: 0.97 -> 0.99 | Helped |
| Use Design2Cloth instead of DrapeNet for cloth reconstruction | Cloth generation on DigitalMe · DigitalMe | CD (×10−2): 0.56 -> 0.12 | Helped |
| Use Design2Cloth instead of DrapeNet for cloth reconstruction | Cloth generation on DigitalMe · DigitalMe | NC: 0.96 -> 0.98 | Helped |
| Use Design2Cloth instead of SMPLicit | Single-image garment reconstruction · CustomHumans | CD (×10−2): 0.40 -> 0.12 | Helped |
| Use Design2Cloth instead of ClothWild | Single-image garment reconstruction · CustomHumans | CD (×10−2): 0.62 -> 0.12 | Helped |
| Use Design2Cloth instead of DrapeNet | Single-image garment reconstruction · CustomHumans | CD (×10−2): 1.14 -> 0.12 | Helped |
| Generate garments with Design2Cloth instead of DrapeNet | Garment realism evaluation · Garment realism user study | average realism score: 4.6 -> 7.2 | Helped |
Surprising findings
- Generated garments were rated more realistic (7.2) than the Cloth3D synthetic ground truth (7.0) in the user study.
Limitations
- Masks are restricted to the frontal part of the garment, so back details are not captured by the conditioning.
- Dataset and code are stated to be released only in future ('will become publicly available'), with no link given.
Experimental setup
Datasets: DigitalMe Cloth3D ClothesNet CustomHumans
Benchmarks: Cloth3D DigitalMe CustomHumans
Metrics: Chamfer Distance Normal Consistency realism score (user study) preference rate (user study)
Baselines: DrapeNet SMPLicit ClothWild
Builds on: SMPLicit DrapeNet MeshUDF PointNet++ Segment Anything
Closest papers by meaning
- gDNA: Towards Generative Detailed Neural Avatars (CVPR 2022 · similarity 0.91)
- FreeCloth: Free-form Generation Enhances Challenging Clothed Human Modeling (CVPR 2025 · similarity 0.89)
- DAGSM: Disentangled Avatar Generation with GS-enhanced Mesh (CVPR 2025 · similarity 0.88)
- ChatGarment: Garment Estimation, Generation and Editing via Large Language Models (CVPR 2025 · similarity 0.88)
- DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images (NeurIPS 2026 · similarity 0.88)
- SCULPT: Shape-Conditioned Unpaired Learning of Pose-dependent Clothed and Textured Human Meshes (CVPR 2024 · 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.