acceptodds
Under review as a conference paper at ICLR 2027

Hash-guided Diffusion: Efficient Generation

Abstract

Physics-based simulation of sewing patterns yields globally correct but overly smooth garment surfaces, and recovering high-frequency detail such as wrinkles and seams with optimization- or diffusion-based refinement typically requires hundreds to thousands of iterations. We present Hash-guided Diffusion (HD), which keeps the simulated surface fixed and generates only a residual displacement field over a frozen UV parameterization, a mapping of the 3D surface onto 2D coordinate charts obtained with Nuvo, a neural UV parameterization. The residual is represented by a multiresolution hash encoding for constant-time lookup of local high-frequency detail, and is generated by a conditional diffusion process with either a discrete (DDPM) or a deterministic probability-flow ODE reverse solver. On a representative stiff garment from GarmentCodeData v2, HD-ODE with 10 reverse steps attains lower panel-level and image-level error than 500-step discrete diffusion (Panel L2 4.72 vs. 5.18 cm). Performance depends strongly on material stiffness, with the best results on stiff fabrics. These results indicate that a geometry-aligned residual domain can substantially shorten diffusion-based surface refinement.

Primary area:
generative models
Keywords:
Diffusion models, Garment surface refinement, Residual diffusion, Neural UV parameterization, Efficient sampling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_U7JVOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a two-stage method, Hash Guided Diffusion. Stage-1: Nuvo constructs a multi-chart YV parameterization of a coarse simulated garment. Stage-2: Diffusion model generates residual displacement fields using hash encoding. DDPM and deterministic ODE are both tested. The results are promising with 10 steps for stiff garments. Howver, this paper lacks clear novelty, enough ablations for claimed benefits, enough evidence for claimed results (rendered qualitative results and inference times). I believe it can be difficult to addrss all concerns within rebuttle period. I suggest rejection.

Critical strengths

The pratical problem for reducing the computational cost of high-frequency surface refinement is well-motivated.

The two-stage design is natrual.

Critical weaknesses
  1. The methodological novelty of this paper is unclear. The proposed pipeline mainly integrate existing methods.
  2. The evidence of design choice is not enough. The authors claimed that the efficient 10 step refinement is from geometry aligned residual representation and hash encoding. However, although the paper introduces baselines without Nuvo and without hash encoding, their numberical result are not included in the main comparision. So it is unclear whether the reduced sampling requirement results from the proposed representation or simply from applying a short-step diffusion solver.
Questions for authors
  1. Why Panel L2 changes when the sewing patterns are fixed?
  2. Could authors provide multi-gartment results, real rendering?
  3. Could authors provide ablation studies regarding w/o Nuvo, w/o hash, direct residual-fitting?
  4. Could authors provide inference times, function evaluations?
1 comment
AuthorsOpenReview ↗

Answers to the Questions one by one

Thank you for the review feedback.

Q1. We will verify the metric implementation and clarify how the metric is computed in the revision.

Q2. We agree that evaluating refinement quality and generalization requires results across multiple garments, with reference renderings and close-up views of the refined surfaces.

Q3. We will verify the implementation and clarify the metric definition and evaluation protocol in the revision.

Q4. As reported in Appendix A.4, the full two-stage optimization takes approximately 2–3 hours per garment on a single NVIDIA H200 GPU.

Reviewer_uXFgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a diffusion-based framework for reconstructing detailed garment surfaces, including fine-grained wrinkles and seams. The proposed method employs a residual diffusion model in the UV domain to reconstruct garment details in a coarse-to-fine manner. In addition, hash encoding is introduced to effectively apply residual diffusion in the UV domain.

Critical strengths
  • Reconstructing fine-grained garment details is an important and challenging problem in the field of garment simulation. The paper addresses a relevant research problem.
Critical weaknesses

Limited technical novelty. The technical novelty of the proposed method is very limited and does not appear sufficient to meet the acceptance standards of ICLR. My main concerns are detailed below.

  • Insufficient differentiation from prior work. The proposed method appears to be essentially an extension of Nuvo with an additional residual diffusion model. The technical contributions beyond the existing framework are not sufficiently demonstrated, making the proposed approach only marginally different from prior work.
  • Unclear motivation for hash encoding. The use of hash encoding appears to be a key component of the proposed method. However, the paper does not sufficiently explain why hash encoding is necessary or what advantages it provides over alternative representations. The technical contribution of this component remains unclear.
  • Lack of conditioning on garment geometry. High-frequency garment details, such as wrinkles, are likely to depend on garment pose and global deformation patterns. However, the proposed model does not explicitly incorporate such information as conditioning signals. This raises concerns about the generalization capability of the proposed method across different garment poses and deformation patterns.
  • Insufficient experimental validation. The experimental evaluation is substantially incomplete. In particular, the paper does not provide any qualitative comparisons, making it difficult to assess the visual quality of the reconstructed garment surfaces or verify the effectiveness of the proposed method.
  • Lack of comparisons with existing methods. The paper does not provide benchmark comparisons against existing approaches. Without such comparisons, it is difficult to objectively evaluate whether the proposed method achieves meaningful improvements over prior work.
Questions for authors

Please refer to the concerns raised in the Weaknesses section, particularly regarding the technical novelty, the motivation for hash encoding, the generalization capability, and the lack of comprehensive experimental evaluations.

1 comment
AuthorsOpenReview ↗

Answers to the Questions one by one

Thank you for the review.

  1. Technical contribution. Nuvo provides the surface parameterization. HD combines a frozen Nuvo atlas, a multiresolution hash grid, and diffusion to model a bounded residual displacement field for per-instance refinement. The contribution is this formulation for combining existing components.
  2. Hash representation. The hash grid provides features at multiple spatial resolutions to represent localized surface details such as wrinkles and seams. Its contribution to refinement quality remains to be isolated through a controlled ablation.
  3. Conditioning and generalization. HD defines the residual relative to the simulated base surface and conditions on material and sewing-pattern information. The model is optimized separately for each garment instance. The reported experiments use a neutral body template, so generalization across garments and body shapes remains untested.
  4. Revision. We will clarify the conditioning inputs, per-instance optimization, and evaluation protocol, and state which generalization claims require further experiments.
Reviewer_NGcROpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Hash-Guided Diffusion (HD), a framework for efficient high-frequency geometric refinement of physics-simulated garments. Rather than generating garments from scratch, the method preserves the coarse garment geometry and predicts surface displacement residuals to recover fine-scale details such as wrinkles and folds. Its main contributions are: (1) a surface-aligned residual representation using a frozen Nuvo UV parameterization; (2) a multiresolution hash-grid encoding for efficient representation of local geometric details; and (3) a conditional diffusion framework supporting both discrete DDPM sampling and deterministic probability-flow ODE sampling, aiming to achieve high-quality refinement with fewer denoising steps. The method is evaluated on GarmentCodeData v2, primarily using a single representative garment, with comparisons across sampling steps and material stiffness settings. The authors report that the 10-step ODE variant achieves lower Panel L2 and L1 errors than 500-step discrete diffusion, suggesting improved sampling efficiency. However, the validation remains limited: no actual qualitative garment refinement results or LPIPS measurements are presented, key component ablations are not quantitatively reported, and the primary Panel L2 metric does not directly measure high-frequency surface detail. Consequently, the experimental evidence is insufficient to establish the claimed refinement quality and efficiency advantages.

Critical strengths

The paper addresses a practical problem in garment modeling: recovering high-frequency geometric details from coarse physics-simulated surfaces. By restricting generation to surface displacement residuals while preserving the underlying garment structure, the proposed formulation offers a reasonable way to decouple global geometry from local detail refinement. This targeted formulation has potential value for high-fidelity garment reconstruction and rendering.

Critical weaknesses

The paper claims to generate high-quality garment wrinkles and fine-scale geometric details, yet provides no qualitative visualizations of actual refinement results. Quantitative evaluation is limited to a single representative garment, and LPIPS, explicitly identified as a primary refinement metric, is not reported. Furthermore, the absence of quantitative ablations for the UV parameterization and hash encoding prevents assessment of their individual contributions. Consequently, the experiments do not adequately establish the effectiveness, generalizability, or necessity of the proposed components. The evaluation relies heavily on Panel L2, a metric measuring 2D sewing-pattern reconstruction error, despite the method explicitly preserving the original sewing-pattern structure and modifying only 3D surface residuals. The reported Panel L2 values nevertheless vary across diffusion configurations, raising serious concerns about the metric's definition and implementation. Additionally, comparison with AIpparel, which primarily addresses sewing-pattern reconstruction rather than surface-detail refinement, does not constitute a task-aligned baseline. These issues substantially undermine the validity of the reported performance improvements and the paper's main conclusions.

Questions for authors
  1. The proposed method optimizes a separate model for each garment instance. What distribution does the diffusion model learn in this setting, and why is diffusion preferable to directly optimizing the hash-based residual field? Clarifying this distinction is important for establishing the necessity of the proposed formulation.
  2. The paper claims effective high-frequency garment refinement but provides no qualitative results, evaluates only one representative garment, and omits LPIPS despite identifying it as a primary metric. What evidence supports the claimed improvement in refinement quality?
2 comments
AuthorsOpenReview ↗

Answers to the Questions one by one

Q1. What does HD learn, and how does it compare with direct fitting?

  1. Per-garment training. HD is optimized separately for each garment instance. The denoiser learns to recover residual targets for that garment from noise-corrupted versions of those targets (Secs. 3–4). We will clarify how the targets are constructed and how residuals are sampled at inference.
  2. Direct-fitting baseline. Whether diffusion improves on direct fitting remains open. A controlled comparison would fit the residual field directly using the same garment, frozen UV atlas, hash representation, supervision, and evaluation protocol.

Q2. What do the current results establish?

We will revise the paper to distinguish the reported measurements from claims about refinement quality and individual components. Establishing refinement quality requires rendered outputs and metrics that measure surface detail across multiple garments. Isolating the contributions of the UV atlas, hash representation, and diffusion objective requires controlled ablations.

Thank you for your response. However, most of my concerns have not been sufficiently addressed. Therefore, I maintain my original recommendation.

Reviewer_vWL9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies garment surface refinement when physics-based simulation produces globally plausible but overly smooth surfaces. It represents residual displacements in a frozen Nuvo UV atlas and combines a multiresolution hash representation with discrete diffusion or a probability-flow ODE. Experiments report results for different reverse-step budgets and material stiffness, with the strongest reported numbers on one stiff garment.

Critical strengths
  1. The paper presents a coherent high-level decomposition between a coarse simulated garment surface and a surface-attached residual detail field.

  2. The reported experiments compare discrete and ODE sampling and examine material stiffness, providing an initial view of the proposed setting’s behavior.

Critical weaknesses
  1. The paper does not provide an in-depth analysis of the problem it aims to solve. As I understand it, the paper focuses on garment surface refinement. However, what causes the need for surface refinement? Have previous works already investigated or addressed this problem?

  2. Why is Hash-Guided Diffusion adopted? What is the motivation for this choice? For each of the associated new components, what is the rationale for introducing it, and what specific problem is it intended to solve?

Questions for authors
  1. What causes the over-smoothing that motivates surface refinement, which prior works have already addressed it, and what precise unresolved gap remains for this paper?

  2. Why is Hash-Guided Diffusion chosen, and what specific problem is each component—the Nuvo atlas, UV residual field, multiresolution hash, DDPM/ODE solver, and L1+LPIPS objective—intended to solve?

1 comment
AuthorsOpenReview ↗

Title: Answers to the Questions one by one

Q1. Why over-smoothing occurs and how this work relates to prior work

  1. Cause of over-smoothing. The pipeline drapes a sewing pattern using a physics-based simulator (Sec. 1, subproblem (ii)). This produces a globally consistent garment surface while omitting details below the simulation resolution.
  2. Prior work. Section 2.1 discusses earlier methods for adding detail to coarse garments. DeepWrinkles (Lähner et al., 2018) and Deep Detail Enhancement (Zhang et al., 2021) predict UV-space normal maps. DiffusedWrinkles (Vidaurre et al., 2024) and D-Garment (Dumoulin et al., 2026) use diffusion to generate UV displacement maps. Li et al. (2025) perform diffusion in sewing-pattern coordinates.
  3. Scope and remaining gap. The methods discussed in Sec. 2.1 train on garment datasets. Our study evaluates refinement on one stiff garment (Table 1 and Sec. 6.2). The current results leave the effects of the representation and the solver unresolved. Separate ablations are needed to isolate those effects.

Q2. Motivation for each design choice

  1. Nuvo atlas (Sec. 4.1). A surface residual field needs stable coordinates. Classical UV unwrapping can fragment irregular meshes (Sec. 2.2). We use a learned, low-distortion atlas with eight charts and freeze it once chart assignments stabilize, so the coordinates remain fixed during residual learning.
  2. UV residual field (Eqs. 2–3). We preserve the coarse mesh as the base surface and predict only a displacement field indexed by 2D surface coordinates. The displacement is bounded to 5% of the mesh scale.
  3. Multiresolution hash grid (Sec. 4.2). Wrinkles and seams contain localized detail at multiple spatial scales. The hash grid provides direct lookup at multiple resolutions (App. B.1, item 2). Including the chart index k in the hash key lets all charts share one hash grid (Eq. 6).
  4. DDPM and probability-flow ODE solvers (Sec. 4.3). The two variants share the forward process and training objective (Eqs. 5 and 15). They use different reverse-time formulations. Table 1 compares the stochastic DDPM and deterministic ODE variants at the reported step counts.
  5. L1 + LPIPS objective (Sec. 4.4 and Fig. 1B). Under squared L2 loss, averaging plausible fold profiles can smooth a crease. In the toy example in Fig. 1B, the L1 median preserves a sharper crease. LPIPS provides a perceptual training signal. Figure 1B illustrates the motivation for the loss and uses synthetic profiles.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.