acceptodds
Withdrawn from ICLR 2027

NOFACE3D: NOise-scheduled Focused Anchoring and Controlled Enhancement for 3D Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent image-to-3D models generate complete, high-quality assets from a single image, with global shape and appearance that are already convincing; what remains is the fine scale, where small, thin, elongated, or detail-rich regions are reconstructed softly and under-resolved. This gap can arise in generators with fixed-resolution spatial grids: the same spatial sampling density is used for a large, smooth, low-curvature surface and for a thin, high-curvature structure that carries most of the perceived detail. To this end, we introduce NOFACE3D, a training-free framework that refines these detail-critical regions while preserving the rest of the asset. Given a generated mesh, its input image, and a refinement box, NOFACE3D extracts the selected submesh, scales it up to fill the generator's input volume, and re-encodes it into the full-resolution latent grid. Conditioned on the image crop, it refines the region through flow-based sampling with a spatially varying noise schedule that anchors boundary latents. The refined region is then integrated back into the original asset, preserving the surrounding geometry and appearance. We implement NOFACE3D on TRELLIS.2 and its sparse voxel representation; all of it operates purely at inference time on the frozen pretrained generator. We evaluate NOFACE3D on PAniC-3D, TexVerse and , a new benchmark of assets with annotated detail-critical regions. On all three, it generates sharper geometry and finer texture detail than the baselines.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Mesh Refinement, 3D Generation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_28T4OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NOFACE3D is a training-free method for refining local details in TRELLIS.2-generated assets. It enlarges a selected mesh region and regenerates it from the corresponding image crop, using a spatial noise schedule to preserve attachment boundaries. The refined region is then merged back into the original asset. Experiments on three datasets show improved visual detail and human preference over the base generator.

Critical strengths
  • The paper proposes a training-free method for improving local details in generated 3D models.
  • The visual examples show clearer faces and small object details after refinement.
  • The user study supports the visual results: 73.3% of all responses preferred the refined models.
Critical weaknesses
  • TRELLIS already supports training-free local editing through RePaint-style masked sampling. The paper should clarify what its spatial noise schedule adds to this approach. A useful baseline would use the same enlarged crop and TRELLIS.2 backbone, but keep boundary latents fixed and use a uniform noise level elsewhere. This would help separate the benefit of local enlargement from that of the proposed schedule.
  • Figure 2 looks blurry, and the box in its fifth panel contains noticeable speckled artifacts. A clearer figure would make the pipeline easier to follow.
Questions for authors
  • Does re-voxelization and remeshing reliably connect the original and refined meshes? Have you measured gaps, disconnected components, or changes outside the selected region in the final output?

  • Can you report geometry metrics within the refined region and near its boundary? Whole-object metrics may hide both local improvements and local errors.

  • Does the final re-voxelization at 1024³ reduce the detail gained through local enlargement? Can you compare the refined region before and after remeshing, especially for thin structures and small gaps?

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

NOFACE3D presents a training-free framework for localized refinement of image-to-3D assets. Given an existing mesh, its conditioning image, and a selected refinement region, the method extracts and rescales the submesh to occupy the generator’s canonical volume, re-encodes it using the pretrained TRELLIS.2 encoders, and regenerates its geometry and texture. A boundary-distance-dependent noise schedule restricts changes near attachment boundaries while allowing stronger regeneration in the interior. The refined region is subsequently transformed back and reintegrated into the original asset. Experiments report improvements in appearance and perceptual quality across PAniC-3D, TexVerse, and the proposed BigDetail benchmark.

Critical strengths

Strengths

  1. A clear and practical motivation. Allocating a fixed spatial representation uniformly across an asset can underrepresent small, detail-rich regions. Reusing the generator’s capacity at a local scale offers an intuitive way to address this limitation without retraining.

  2. Explicit consideration of attachment boundaries. The spatial release schedule and repeated anchoring acknowledge that independently regenerating a region can introduce geometric and appearance discontinuities. The proposed pipeline makes a meaningful effort to balance local enhancement with preservation.

  3. Encouraging perceptual improvements. The evaluation includes reference-based metrics, qualitative comparisons, model-based judgments, and a human preference study. Together, these provide evidence that the refinement can improve the perceived quality of generated assets.

Critical weaknesses

W1. The novelty relative to existing decomposition-and-generation pipelines is insufficiently established. The high-level strategy of generating selected regions separately and assembling them into a larger 3D representation has substantial precedent. In particular, SAM 3D Objects reconstructs mask-conditioned objects together with their geometry, texture, and spatial layout. Although object-level scene assembly differs from replacing a connected subregion of an existing mesh, this related work makes it important to identify precisely what NOFACE3D contributes beyond localized generation and placement. The paper should discuss this distinction explicitly and include a task-matched comparison where feasible. For example, a segmentation-guided local reconstruction pipeline could regenerate the same selected region and place it back into the source asset, with alignment and reintegration procedures documented. Boundary quality, local fidelity, preservation, and computational cost would then help establish the additional value of NOFACE3D. The capacity-allocation argument is also relevant to scenes or multi-object assets, where small objects receive limited representation capacity. A scene-level experiment would help establish the broader usefulness of the approach, although the authors should distinguish independent-object replacement from connected-surface refinement. At present, the conceptual contribution appears incremental, and the necessity of the particular sampling mechanism is not sufficiently isolated.

W2. Robustness to local reparameterization and the loss of global context needs stronger investigation. Extracting, rescaling, and re-encoding a component changes its coordinate frame, spatial discretization, and available context. Consequently, its local latent representation need not match the representation it had within the complete asset. This difference is not inherently problematic, but it raises questions about whether the regenerated geometry remains compatible with its surroundings beyond the explicitly anchored band. The existing ablation removes boundary anchoring and spatial scheduling together and primarily provides qualitative evidence. It therefore does not determine whether the proposed schedule specifically addresses these issues, or whether a simpler hard boundary mask with uniform interior noise would suffice. I would encourage controlled experiments varying crop size, crop position, retained context, and the number of attachment boundaries. Particularly useful measurements would include local reconstruction error, boundary position and normal discontinuities, changes outside the selected region, and consistency across slightly perturbed boxes. These experiments would clarify whether the method is robust to reasonable region selections rather than dependent on carefully chosen crops.

W3. Improvement of unseen and geometrically complex surfaces remains insufficiently demonstrated. Rescaling a region increases its effective spatial resolution, but it does not introduce new observations. The refinement remains conditioned on a single image crop, which may provide little information about side surfaces, rear surfaces, cavities, or occluded connections. The method can synthesize such geometry through its pretrained prior, but plausible additional detail should be distinguished from faithful recovery. This distinction matters for complex objects, where a sharper frontal appearance can coexist with incorrect hidden geometry. The reported whole-asset metrics do not adequately isolate these cases, and preference-based evaluation alone cannot establish geometric accuracy. The authors should evaluate the selected region separately across visible and occluded surfaces, using reference meshes where available. Region-specific geometry metrics, side and rear views, sectional visualizations, and examples involving cavities or thin connected structures would be informative. An optional multiview-conditioning experiment could further establish whether the observed limitations arise from insufficient visual evidence or from the refinement procedure itself.

Questions for authors

Please refer to weakness.

Reviewer_2j5cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for local 3D asset refinement based on pre-trained generative models that requires no additional training. Given an input mesh and a selected refinement region, the method crops and scales the corresponding submesh to better utilize the generator’s latent resolution. Local refinement is then performed using spatially varying noise and boundary re-anchoring. Finally, the refined regions are mapped back to the original coordinate system and reintegrated into the base asset, resulting in a fully refined asset. The main technical contribution of this paper is a boundary-aware sampling strategy: latent variables near connected boundaries more strongly preserve consistency with the original asset, while regions farther from the boundaries are released earlier to allow for stronger regeneration. The authors conducted evaluations on multiple datasets and benchmarks, providing qualitative comparisons and limited ablation studies of the refinement strategy. According to the results reported in the paper, the complete framework is generally capable of improving the quality of local details compared to the original asset and several competing refinement methods.

Critical strengths
  1. This paper is written clearly, and the implementation details are fairly transparent. The paper is generally well-written, and the methodological workflow is relatively easy to understand. The supplementary materials provide further details on the implementation setup and design, which help enhance the transparency and reproducibility of the method.

  2. Extensive comparisons were conducted with existing methods. The paper compares the approach with a number of local refinement baselines across multiple datasets and evaluation metrics, thereby demonstrating to some extent that the complete pipeline is competitive in terms of practical performance.

Critical weaknesses
  1. The core refinement strategy in this paper is primarily based on the adaptation and combination of existing techniques, including pre-trained generative priors, distance-based spatial scheduling, flow-based sampling, and re-anchoring replacement. Although the overall workflow is practical, the paper currently fails to adequately explain what substantive algorithmic innovations it proposes compared to simply applying these existing techniques to local 3D refinement.

  2. The proposed core components have not been fully validated. While the paper provides a comprehensive comparison of the proposed method with other approaches, it fails to sufficiently decompose and validate the core design through ablation experiments. The only ablation experiments conducted so far are qualitative and remove multiple components simultaneously, making it difficult to determine whether the final improvement stems from a single module or from their combination. Given that these designs constitute the paper’s primary technical contributions, this omission directly affects the assessment of the method’s effectiveness and innovativeness.

Questions for authors
  1. The current ablation experiments are insufficient, making it difficult to determine which specific component of the refinement strategy is responsible for the final performance improvement. Could the authors validate the results of individual components, such as boundary anchoring and the spatially varying noise schedule, as well as different designs within those components?

  2. The computational advantages of the training-free design have not been fully validated. The paper provides the absolute runtime of the proposed pipeline and notes that its sampling cost is roughly equivalent to that of the underlying generator. However, there is still a lack of a systematic end-to-end cost comparison with other refinement methods under uniform hardware conditions. Since the training-free approach is presented as one of the paper’s strengths, reporting a comparison of overall overhead, including preprocessing, would make this advantage more convincing.

  3. Since the band widths (g,r) are empirically selected and τmax⁡\tau_{\max} directly controls the refinement strength, a brief sensitivity analysis would help assess the robustness of these choices.

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

Related papers

Discussion (0)

Sign in to comment.