OmniFabric: Coherent UV Space Texture Synthesis for 3D Garment Reconstruction

• Current 3D texturing methods struggle with garment folds, self-occlusions, and pose variation, often baking lighting and geometry-induced wrinkles into textures during projection.These distortions and transient artifacts undermine the intrinsic albedo needed for downstream use.Source: §1 · p. 2
• FabricDiffusion produces simulation-ready sewing-pattern textures but misses global layout, limiting reconstruction of asymmetric designs, large prints, and irregular patterns.The paper argues that these designs require context from the complete reference image rather than isolated patches.Source: §1 · p. 2
• OmniFabric aligns a reposed reference garment to a canonical mesh, then uses video-model multi-view observations to initialize globally coherent textures on unwrapped sewing patterns.The initialization is complete but coarse and can contain geometric distortions and physics-induced wrinkles.Source: §1 · p. 2
• An automated synthetic-data engine trains a sewing-pattern-space Diffusion Transformer to rectify distortions and remove transient illumination, producing normalized textures.The listed contributions also include using a canonical 3D mesh as a spatial anchor for globally aligned texture synthesis.Source: §1 · p. 2
• Grounding synthesis in sewing-pattern UV space is presented as a way to maintain global coherence while staying aligned with practical garment construction.Source: §1 · p. 2

2.1 3D Garment Reconstruction

• Garment reconstruction has progressed from implicit surfaces to generative and sewing-pattern methods, but normalized, artifact-free textures for sewing panels remain an open challenge.Some simulation-ready approaches use computationally intensive per-asset optimization, while sewing-pattern methods predict or generate garment structures.Source: §2.1 · pp. 2–3

2.2 Generative 3D Texturing

• General-purpose projection-based texturing methods can generate coherent global structures, yet natural-photography priors entangle illumination with appearance and complex topology introduces distortion and blur.The paper also describes earlier optimization-based methods as prone to generic textures lacking high-frequency detail.Source: §2.2 · p. 3
• FabricDiffusion extracts tileable, distortion-free material patches without lighting artifacts, but local tiling misses graphic placement and non-uniform patterns across UV islands.Source: §2.2 · p. 3
• OmniFabric shifts from local extraction to global canonical-UV synthesis, while prior UV refinement methods inpaint occlusions but do not correct distortion or remove baked-in artifacts.Figure 2 illustrates the stated limitations of baseline garment texturing methods.Source: §2.2 · p. 3, Figure 2 · p. 3

3.1 Problem Definition

• Given a reference image and an estimated sewing-pattern-parameterized mesh, the goal is a global normalized texture map free of distortion, view-dependent illumination, cast shadows, and geometry-induced wrinkles.The texture is defined as a conditional distribution mapping problem.Source: §3.1 · p. 3

3.2 Coarse Texture Map Initialization

• The coarse initialization uses pose-aware alignment and multi-view generative priors to establish pixel-wise correspondences and cover the sewing-pattern UV space.The two-stage overview describes this initialization as globally aligned but coarse.Source: §3.2 · p. 4, Figure 3 · p. 4

3.2.1 Pose aware texture alignment.

• Reposing the reference image to match the rest-pose mesh silhouette enables direct projection onto the mesh and sewing patterns, producing a textured frontal render.Source: §3.2.1 · p. 4

3.2.2 Consistent multi-view synthesis.

• A video model generates a 360-degree garment sequence conditioned on the frontal render; four orthogonal views are projected onto the mesh and mapped into UV space.Source: §3.2.2 · p. 4

3.3 UV Space Texture Refinement

• The refinement model converts the artifact-ridden coarse map into a normalized texture through holistic UV synthesis with multimodal conditioning.The coarse map can contain shadows, physical wrinkles, unpainted areas, and spatial distortions.Source: §3.3 · p. 4, §3.3.1 · p. 4

3.3.1 Training objective of multi-condition diffusion.

• The diffusion model conditions on a coarse texture map, textured frontal rendering, and 3D position map to learn spatial connectivity across separate UV panels.The position map is projected onto the 2D panel layout and helps ensure seamless textures across fragmented UV islands.Source: §3.3.1 · pp. 4–5
• A DiT processes condition and noisy-latent tokens jointly, using shared positional encodings across panel-layout maps while retaining distinct coordinates for the frontal rendering.The model is fine-tuned on synthetic textured sewing-pattern data with LoRA.Source: §3.3.2 · p. 5

4 Synthetic Training Data Creation

• The data engine targets realistic garment appearances while pairing simulated texture initializations with normalized ground-truth maps for supervised training.The construction addresses a dataset bottleneck: available texture sources often provide only simple, repetitive patterns.Source: §4 · p. 5, §4.1 · p. 5

4.1 Textured Sewing Pattern Creation

• Fashion-image references guide generated texture designs: an LLM describes layout, semantic elements, and colors, then a vision-language model synthesizes diverse textures over sampled sewing patterns.The generated textures need to be plausible and diverse, not exact reconstructions of the example images.Source: §4.1 · p. 5

4.2 Training Pairs Construction

• Training pairs use a TPS-distorted initial texture overlaid on sewing patterns, with the original undistorted texture as ground truth.After simulation, frontal renderings and positional maps are generated, and projections produce coarse maps containing physics-induced artifacts and deformations.Source: §4.2 · pp. 5–6

5.1 Setup

• Evaluation uses 30K synthetic textured sewing patterns from 3K garment samples, standard perceptual, structural, and semantic metrics, and in-the-wild images assessed qualitatively.The in-the-wild set combines DeepFashion images with clothing-design images collected from the web; quantitative evaluation uses a held-out synthetic test set.Source: §5.1.1 · p. 6, §5.1.2 · p. 6

5.2.2 Simulation results.

• The generated garments show seamless textures and convincing appearance under varied poses and lighting, including extreme dynamics.The simulation uses normalized RGB base-color as albedo; other material parameters and normal maps use CLO defaults.Source: §5.2.2 · pp. 6–7, §5.2.1 · p. 6, Figure 5 · p. 7

5.2.3 Comparison with state-of-the-art methods.

• For complex in-the-wild textures, OmniFabric preserves coherent details, while Hunyuan3D and Paint3D lose high-frequency detail and FabricDiffusion lacks global coherence.The qualitative examples include asymmetric textures and logos.Source: §5.2.3 · pp. 7–8, Figure 6 · p. 8

5.3.1 Synthetic dataset.

• On the synthetic benchmark, OmniFabric excels across all reported metrics, particularly pixel and perceptual quality, global structural coherence, and projection-artifact removal.Baselines receive identical frontal renderings and rest-posed meshes; their textures are projected into sewing-pattern UV space for comparison.Source: §5.3.1 · p. 8, Table 1 · p. 8

5.3.2 Real-world data.

• In a study of 13 participants judging 10 real-world inputs, OmniFabric ranks best across overall quality, fidelity, back-view plausibility, and global texture coherence.Overall-quality ranking uses a 1-to-4 scale, lower is better; the other criteria use 1-to-5 scales, higher is better.Source: §5.3.2 · p. 8, Table 2 · p. 8

5.4 Reproducibility via Open-Source Models

• A fully open-source alternative achieves performance similar to the default closed-source-model setup under the same experiment settings.The alternative uses FLUX.2 and Ministral 3 in place of the Gemini models.Source: §5.4 · pp. 8–9, Table 3 · p. 9

5.5 Error-Accumulation Analysis

• Replacing predicted intermediate outputs with ground truth exposes accumulated errors that texture normalization can remove, but inaccurate multi-views can still cause mismatched textures and incorrect input geometry remains uncorrected.The refinement model reduces distortions, wrinkles, shadows, and projection artifacts given a sewing pattern.Source: §5.5 · p. 9, Table 4 · p. 9, §5.7 · p. 9

5.6 Ablation Studies and Analyses

• Ablations show that the coarse texture map and frontal rendering provide important guidance, while removing positional conditioning misplaces details; distortion-augmented training supports distortion correction.Without the coarse map, the model generates textures that look similar rather than pixel-aligned; without the frontal rendering, results are inferior.Source: §5.6.1 · p. 9, §5.6.2 · p. 9, §5.6.3 · p. 9

6 Conclusion and Discussion

• OmniFabric produces normalized, globally coherent textures for 3D garments and separates intrinsic albedo from baked-in shadows and wrinkles.The study focuses on sewing patterns within the GarmentCode framework.Source: §6 · p. 9
• The current study predicts albedo; jointly predicting roughness, metallic, and normal maps is proposed as a future extension.The paper states this extension could further enhance photorealism under diverse environmental lighting.Source: §6 · p. 9
• Generated garments can be simulated under varying human poses and lighting conditions with convincing results.The simulation examples show garments generated by the method.Source: Figure 9 · p. 10

A Details of Dataset Construction

• The automated data engine generates diverse, high-fidelity textured sewing patterns to support dataset construction at scale.The authors identify the scarcity of complex, high-resolution textures as a need addressed by the engine.Source: A Details of Dataset Construction · p. 12

A.1 Reordering Sewing Pattern for Texturing

• Pattern pieces are reordered into functional groups so textures remain continuous within groups but break at structural seams, matching garment-construction needs.The strategy merges related panels, such as front and back pieces of a sleeve.Source: §A.1 · p. 12

B Additional Implementation Details

• The multi-view projection uses four orthogonal frames from a canonical A-pose rotation video and a rest-posed mesh reconstructed from predicted patterns.The views correspond to front, back, left, and right.Source: B Additional Implementation Details · p. 12

B.1 Non-overlapping Texture Projection

• Front and back views are projected first; side views fill only untextured regions, and normalization corrects remaining seam artifacts while retaining high-frequency detail.This non-overlapping strategy uses frames from a structurally consistent video prior.Source: §B.1 · p. 12

B.2 Prompt Configurations

• The prompt configurations request a textured frontal view matching the reference garment, then a continuous 360-degree rotation video.The frontal-view prompt asks to preserve the target garment’s style, size, and pose.Source: §B.2 · p. 12

B.3 Training Details

• Paint3D’s position encoder, Hunyuan3D’s multi-view image generator, and FabricDiffusion’s texture generator were fine-tuned from pretrained weights.FabricDiffusion’s texture generator was trained using paired textile data.Source: §B.3 · pp. 12–13
• Figure S1 presents a synthesized textured sewing pattern and its simulated 3D garment for each synthetic-data example.Source: Figure S1 · p. 13

C Additional Results and Analyses

• Figure S2 shows OmniFabric integrating with another sewing-pattern prediction method.Its simulation results are described as demonstrating seamless integration.Source: Figure S2 · p. 13

C.1 Robustness to Structural Variations in Sewing Patterns

• Across garment styles, normalization removes projection-induced distortions and physics-based wrinkles from coarse texture maps.The evaluation also reports robustness when simulated geometry deviates from the input silhouette; texture transfer aligns image textures with the rendered silhouette to preserve global coherence.Source: §C.1 · p. 13

C.2 Compatibility with Other Sewing Pattern Prediction Method

• With Alpparel [Nakayama et al. 2025] sewing patterns, OmniFabric generates garments with aligned geometry and high-frequency details.The evaluation supports compatibility beyond the ChatGarment pattern predictor used in the main paper.Source: §C.2 · p. 13
• Imperfect pattern prediction can leave the simulated garment geometry mismatched to the input, although OmniFabric maintains global texture consistency.The framework is described as agnostic to the reconstruction method and compatible with future pattern-prediction advances.Source: §C.2 · pp. 13–15
• Qualitative results show normalized sewing patterns with projection distortions and geometry-induced wrinkles removed.Source: Figure S3 · p. 14

D Limitations and Future Work

• OmniFabric removes transient lighting artifacts but does not explicitly disentangle albedo from shading in a principled, physics-based way.Source: D Limitations and Future Work · p. 15
• Generative priors can hallucinate unseen or occluded regions; multiple frames or video could reduce these hallucinations.Source: D Limitations and Future Work · p. 15
• Highly reflective leather and metallic fabrics remain challenging; the authors suggest jointly predicting complete PBR maps, including roughness and normal maps.Source: D Limitations and Future Work · p. 15