Research map / 3D vision / Face and shape

DressWild: Feed-Forward Pose-Agnostic Garment Sewing Pattern Generation from In-the-Wild Images

NeurIPS 2026 · 3D vision › Face and shape · approach: Clothed human modeling · 0 citations (OpenAlex)

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 DressWild, a feed-forward pipeline that predicts 2D sewing patterns and 3D garments from a single in-the-wild image by VLM-based canonical T-pose normalization and fusion of pose, appearance, and canonical features.

Problem

Sewing pattern generation from in-the-wild images with diverse poses and viewpoints; existing feed-forward methods are tied to canonical poses and optimization-based methods are slow.

Approach

Use a VLM (Nano Banana Pro) to synthesize a canonical front-facing T-pose garment image, extract appearance, canonical (Hunyuan3D latent) and pose (SAM3D-Body) features, fuse them with a transformer encoder, and autoregressively decode parametric sewing pattern parameters.

Key results

• Panel accuracy 94.35% and edge accuracy 85.41% vs. 28.81%/34.56% for SewFormer and 25.99%/29.05% for NeuralTailor.
• Shape L2 error 6.22 vs. 22.94 (SewFormer) and 23.65 (NeuralTailor); Chamfer distance 0.01899 vs. 0.02797 (SewFormer).
• Removing canonical front-view features drops panel accuracy from 94.35% to 91.18% and raises shape error from 6.22 to 12.11.

Surprising findings

• SewFormer achieves a lower F-shape error (4.53) than DressWild (5.07), which the authors attribute to that metric ignoring wrong panels.

Limitations

• Relies on proprietary VLM (Nano Banana Pro) for canonicalization and data augmentation, with synthesis instability noted.
• Lower F-shape error than SewFormer, indicating weaker shape fidelity on some metrics.
• Multi-layer garments handled by VLM decomposition without a joint model.

Experimental setup

Datasets: GarmentCodeData-derived dataset (20,000+ designs, 19 base types) VLM-generated multi-pose multi-view images

Benchmarks: in-the-wild garment test split (7:2:1 of 25031 samples)

Metrics: panel accuracy edge accuracy shape L2 error F-shape L2 error Chamfer distance

Baselines: NeuralTailor SewFormer

Builds on: Hunyuan3D-2 SAM3D-Body HybridGL GarmentCodeData CIPC

Closest papers by meaning

• SwiftTailor: Efficient 3D Garment Generation with Geometry Image Representation (CVPR 2026 · similarity 0.91)
• ChatGarment: Garment Estimation, Generation and Editing via Large Language Models (CVPR 2025 · similarity 0.90)
• Reconstruction of Manipulated Garment with Guided Deformation Prior (NeurIPS 2024 · similarity 0.89)
• FreeCloth: Free-form Generation Enhances Challenging Clothed Human Modeling (CVPR 2025 · similarity 0.88)
• Garment Recovery with Shape and Deformation Priors (CVPR 2024 · similarity 0.88)
• Design2Cloth: 3D Cloth Generation from 2D Masks (CVPR 2024 · similarity 0.88)