acceptodds
Withdrawn from ICLR 2027

Dual-Geometry Carrier Atlas Transport for Efficient and Consistent Single-Reference 3D Human Garment Editing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing 3D human garment editing methods typically follow a paradigm that repeatedly performs diffusion inference on multi-view images as references and updates 3D Gaussians accordingly. Repeatedly and independently applying diffusion models leads to substantial computational cost and cross-view inconsistency. In this paper, we propose Dual-geometry Carrier Atlas Transport (DCAT), which edits a single reference view and propagates the appearance to all views for 3DGS supervision. DCAT anchors the edited garment to a garment-adaptive carrier and transports its appearance through a shared UV atlas with side-view feedback, thereby coupling geometry adaptation with cross-view propagation. Specifically, Garment-Adaptive Dual-geometry Carrier deforms an SMPL-X shell according to the edited silhouette for normal garments and constructs an independent skirt proxy for skirt-like garments, thus reducing silhouette mismatch and texture distortion. Built on the carrier, UV Atlas Transport with Side-View Feedback bakes the reference appearance into a shared atlas and propagates it to the extreme side views. Boundary-locked diffusion then complements the side-view regions unobserved in the reference view, after which the results are harmonized and fed back into the atlas for propagation to the remaining views. Extensive experiments on diverse garments and editing instructions demonstrate that DCAT significantly improves editing efficiency and cross-view consistency while maintaining high visual quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3DGS, Gaussian editor, Human garment editor

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_rgU8OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Dual-geometry Carrier Atlas Transport (DCAT), a single-reference framework for efficient and consistent text-driven 3D human garment editing. DCAT combines a garment-adaptive dual-geometry carrier, comprising a silhouette-aligned SMPL-X shell and an independent skirt proxy, with UV Atlas Transport and Side-View Feedback to propagate a single reference edit across views, improving editing efficiency and cross-view consistency while maintaining high visual quality.

Critical strengths
  1. Important task. Existing 3D human garment editing methods update 3D Gaussians through independent per-view diffusion editing, resulting in high computational costs and cross-view inconsistencies. DCAT propagates edited garment appearance across views using a garment-adaptive carrier and a shared UV atlas with side-view feedback.
Critical weaknesses
  1. Small evaluation scale. The evaluation covers only two human subjects, providing insufficient evidence of generalization across diverse body shapes, poses, and garment types.
  2. Limited garment-structure coverage. The skirt-specific proxy addresses one type of structural mismatch, but the applicability of the dual-carrier design to other challenging garments, such as capes, open coats, and layered clothing, remains unclear.
Questions for authors
  1. Figure labels. Please correct the LM/RM definitions in the lower-left corner of Figure 2.
  2. Evaluation protocol. What reference images are used for PSNR/SSIM/LPIPS in Table 1? Are test views excluded from atlas construction and side-view completion?
  3. Evaluation scale. Could the authors evaluate beyond two subjects to demonstrate generalization across body shapes, poses, and garment types?
  4. Garment-structure coverage. Can DCAT handle capes, open coats, and layered clothing without additional proxies? Please provide examples or clarify its limitations.
Reviewer_6A37OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DCAT, a single-reference 3D human garment editing framework for 3D Gaussian Splatting (3DGS) scenes. To avoid the high computational cost and texture flickering of multi-view SDS, the method edits a single frontal view using a 2D diffusion model, projects this edit onto a coarse 3D proxy mesh, and completes occluded side views via Flux2 before baking the consolidated texture back into an UV map. This baked texture is then rendered from 20 novel viewpoints to supervise the 3DGS scene update.

Critical strengths

Formulating 3D human garment editing as a single-view 2D edit followed by geometry-guided UV transport is a sensible way to bypass the multi-view inconsistencies of iterative 2D diffusion editing. The paper shows decent editing results compared to baselines.

Critical weaknesses
  1. My primary concern is the significant lack of novelty. The entire "texture transport" pipeline is conceptually identical to established 2023–2024 mesh retexturing frameworks like TEXTure (SIGGRAPH 2023), Text2Tex (ICCV 2023), and Paint3D (CVPR 2024). DCAT simply projects a 2D frontal edit into a UV map, render side views to do inpainting plus classical Poisson/Laplacian harmonization. I struggle to find any fundamental algorithmic contribution to texture modeling here.

  2. The method's geometry pipeline is fundamentally flawed. Shifting SMPL-X vertices along surface normals to match a single 2D frontal silhouette (Eq. 3) is a highly under-constrained inverse problem. Because a frontal mask provides zero depth (Z-axis) constraint, the proxy is mathematically forced to assume a perfectly smooth, uniform "sausage-like" inflation (or a rigid cone for skirts). How can this representation faithfully represent diverse and complicated garment geometries, including folds, pleats, or collars? Any high-frequency details would be baked entirely flat onto the UV map.

  3. The experimental comparison is weak. The baselines are primarily early 2024 multi-view editors (GaussianEditor, DGE, EditSplat). The authors omit critical contemporary baselines that are highly relevant to their claims. For example, Texture-GS (ECCV 2024) and UVGS (CVPR 2025), which similarly map 3DGS into a 2D UV space to resolve view inconsistency. Furthermore, the related work section fails to discuss single-view 3D garment editing and multi-view completion papers like GarmentCrafter (3DV 2026) and Garment3DGen (3DV 2025), as well as hybrid mesh-3DGS garment representations like Gaussian Garments (3DV 2025).

Questions for authors
  1. How does the single-view silhouette optimization handle subjects with complex poses, such as crossed arms, where the arms overlap and occlude the garment's frontal boundary?

  2. Can you compare your method against Texture-GS or UVGS to justify your specific choice of normal-inflated SMPL-X carriers over direct 3DGS-to-UV projection?

Reviewer_AFdkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DCAT edits an existing human 3D Gaussian scene using a text instruction and a fitted SMPL-X model. It edits one reference image and transfers the appearance through a shared UV atlas. A silhouette-aligned SMPL-X shell supports general garments, while a separate elliptical skirt proxy avoids the leg topology of the body model. Two side views are completed with boundary-locked diffusion, color-harmonized, and baked back into the atlas; propagated views then supervise Gaussian optimization. Experiments compare several 3D editors and virtual try-on methods on two human scenes, with IP2P- and FLUX.1-based reference editing and qualitative component ablations.

Critical strengths
  1. The method addresses a concrete cross-view editing failure through explicit correspondences. A common atlas makes the propagation of garment patterns inspectable, while the adapted shell and skirt proxy address silhouette and topology mismatches that a body-tight carrier cannot handle. Preserving observed texels during side completion is a sensible way to prevent diffusion from repeatedly changing already established appearance. The figures and sampled media show coherent garment colors and patterns over the supplied camera changes, supporting the practical value of this pipeline.

  2. The reported efficiency improvement is substantial in the tested setting. Table 1 reports an average of 83 seconds for DCAT-IP2P, versus 232 for DGE and 436 for GaussianEditor; the FLUX.1 variant takes 92 seconds versus Omni-3DEdit's 282. The qualitative ablations show recognizable failure modes when shell adaptation, skirt support, or side completion is removed. These results support the engineering premise that a few appearance-generation steps plus geometric transport can reduce repeated multi-view editing cost.

Critical weaknesses
  1. Independent validation is too limited to support the paper's broad quality and generality claims. Section 4.1 uses edited and propagated images as pseudo-ground-truth targets for held-out-view PSNR/SSIM/LPIPS. This can measure agreement with a chosen synthetic supervision set, but does not independently establish correct garment geometry or editing quality; it is also unclear whether targets are shared across methods or generated separately. Section 4.3 and Appendix A limit the study to two reconstructed people, with 15 IP2P instructions and four FLUX.1 instructions. Multiple garment prompts on the same two scenes do not establish robustness to varied body poses, segmentation quality, garment geometry, or capture conditions. CLIP scores and selected visual examples provide some complementary evidence, but do not resolve this limitation.

  2. The comparisons do not isolate the proposed transport mechanism from the additional completion model and associated inference budget. Section 3.2 uses FLUX.2-klein-4B for the side completions, including in the pipeline whose reference editor is IP2P. The IP2P baselines are described as using IP2P, without an equivalent completion prior. Thus the quality advantage cannot be attributed solely to the dual carrier and atlas feedback. The ablation removes completion altogether, rather than comparing alternative propagation under the same editor, side-generation model, and diffusion budget. The headline single-reference formulation should also be understood as one full reference edit plus additional side-view diffusion, not one diffusion invocation. This distinction matters for both the scientific contribution and the efficiency comparison.

Questions for authors
  1. Exactly how are the pseudo-ground-truth images constructed for each method? Are all methods evaluated against the same target images, or against their own edited/propagated images? Which views enter atlas construction, side completion, 3DGS fitting, and testing? Please clarify what independent conclusion about editing quality or consistency the reported PSNR/SSIM/LPIPS can support, beyond fitting the generated supervision.

  2. Does DCAT-IP2P use FLUX.2-klein-4B for both side completions, and do the baselines receive an equivalent generative prior? Please provide the exact number of model calls and denoising steps, identify whether the reported time includes segmentation, carrier fitting, UV processing, completion, and 3DGS updating, and indicate any existing comparison with editor and diffusion budget held fixed.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.