acceptodds
Withdrawn from ICLR 2027

SUPRA-Vox: Native Voxel Prompts for Super-Aligned 3D Asset Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image-to-3D models can generate visually compelling assets, but often struggle to faithfully reproduce the geometry visible in the input view. In multi-stage 3D generation, missing parts or misplaced surfaces in the initial structure can persist in the final shape, even after later stages add fine details. We propose SUPRA-Vox, a surface-guided framework for sparse voxel generation toward what you see is what you get geometry. SUPRA-Vox encodes the visible surface reconstructed by a geometry foundation model intoNative Voxel Prompts using the generator's pretrained variational autoencoder (VAE). These prompts remain available as separate conditions throughout denoising. Together with a visual hull mask, they guide the inpainting transformer as it jointly updates the complete voxel structure. The resulting structure is passed to a frozen downstream generator for detailed shape synthesis. We train SUPRA-Vox on a curated collection of high-quality 3D assets and evaluate it on two complex-geometry benchmarks we construct, Complex-Geo and DTC-Geo. Experiments on these benchmarks demonstrate state-of-the-art visible-surface reconstruction performance among the evaluated methods and improved overall geometric quality over native Pixal3D, demonstrating that better early structure benefits final geometry.

Primary area:
optimization
Keywords:
3D Generation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_79TcOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper propose to use point maps from Moge3 as the conditioning signal so that the visible surface is align with the image and use generative prior from Pixal3D to fill in the hidden parts. Their contribution is to incorperate point maps with Pixal3D through new conditionings.

Critical strengths

None

Critical weaknesses
  • The paper is poorly written. There are so many notation without clear explaination and notation definition.
  • The novelty is very limited. Just unproject the point maps back to the voxel space and recast Trellis2 for inpainting
  • There is no explaination why the authors propose concatenation architecture. There are so many design choices that you could condition it on, for example, through cross attention [1] or injecting it directly in each voxel [2].
  • The paper constructs their own datasets which might be unfair to others. The author should have provided the comparisons from the publicly-available dataset.

[1] RecGen 3D Multi-Object Scene Reconstruction from Sparse Observations

[2] GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction

Questions for authors
  • If we were to generate transparent objects, how can this approach help as Moge3 will give the point maps only on, for example, the bottle surface but not fish inside the bottle. I don't see any transparent reconstruction in the paper and supplementary.
  • How do you get camera pose TT? This is single image reconstruction so camera pose doesn't have any meaning here? TT could be identity matrix.
  • Why do you need training for this? Can you just doing outpainting from Trellis2 using point maps to constraint the generation?
  • How do you render normal maps? Which camera angles do you use?
Reviewer_BhUUOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

SUPRA-Vox augments Pixal3D with explicit visible-surface conditioning for single-image 3D asset generation. A monocular geometry estimator predicts a partial surface, which is calibrated using native Pixal3D sparse-structure predictions, voxelized, and encoded with the frozen structure VAE. The resulting latent and a depth-derived completion mask are supplied throughout denoising. The sparse-structure generator is fine-tuned, while the downstream shape generator remains frozen. Experiments on two curated benchmarks report improvements in visible-surface fidelity and aligned whole-shape geometry over native Pixal3D.

Critical strengths
  1. Pixel-aligned image conditioning does not explicitly determine where the visible surface lies along each viewing ray. Introducing estimated surface geometry provides a complementary signal for preserving visible structures.
  2. The evaluation includes two benchmarks, ground-truth depth and structure controls, alternative geometry estimators, conditioning and calibration ablations, and an informative failure case. Several visible-surface metrics improve consistently across both benchmarks.
Critical weaknesses
  1. The method relies on native Pixal3D predictions for calibration, reuses its structure VAE and image-conditioning mechanism, fine-tunes its sparse-structure transformer, and directly uses its frozen downstream generator. The main architectural modification is additional conditioning channels under the original flow-matching objective. While this is a reasonable extension, the distinctive benefit of persistent native-latent conditioning over simpler geometric conditioning or existing partial-geometry inpainting approaches, such as Points-to-3D, is not convincingly established.
  2. The main comparison is against native Pixal3D, whereas SUPRA-Vox receives additional fine-tuning on HQ100k. A Pixal3D baseline fine-tuned on the same data with a matched training budget is missing. Furthermore, the depth-zero ablation removes both the prompt and mask from an already trained conditional model without retraining, while retaining MoGe-derived camera parameters. This demonstrates reliance on the supplied conditions, but does not establish their benefit over a properly trained, matched baseline.
  3. Replacing Pixal3D-based calibration with surface-based calibration substantially reduces F@2% in the reported ablation. Training also uses ground-truth depth, whereas inference uses predicted geometry. The failure case demonstrates that incorrect surface estimates can propagate into the final mesh. The paper does not sufficiently establish whether its conditioning mechanism handles such errors better than alternatives, or quantify the end-to-end overhead of the additional geometry estimation and native Pixal3D structure pass.
Questions for authors
  1. How much improvement remains over an RGB-only Pixal3D baseline fine-tuned on HQ100k with matched training budget, camera handling, and sampling settings?
  2. With the backbone, data, and predicted geometry fixed, how does persistent native-latent conditioning compare with simpler depth/voxel conditioning or latent-mixing/inpainting alternatives? Can the prompt and mask contributions be isolated?
  3. How sensitive is the method to depth errors, missing surface regions, and camera inaccuracies? Does persistent conditioning improve robustness relative to alternative observation-injection strategies?
Reviewer_6oCFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SUPRA-Vox, a surface-guided framework for image-to-3D generation that improves geometric alignment with the input image. The method uses a pretrained geometry estimation model to reconstruct the visible surface, which is voxelized and encoded into the sparse-structure latent space of Pixal3D. These Native Voxel Prompts, together with a visual hull mask, condition the sparse-structure generation transformer throughout denoising. The generated sparse structure is subsequently processed by a frozen downstream shape generator.

The authors construct two benchmarks, Complex-Geo and DTC-Geo, and demonstrate improvements in visible-surface reconstruction metrics over existing methods, along with modest improvements in whole-mesh geometry compared with native Pixal3D.

Critical strengths
  • Effective integration of geometric priors into 3D generation. Incorporating explicit visible-surface observations into the native voxel representation is a reasonable approach to improving geometric consistency. The proposed method leverages existing pretrained components without modifying the downstream shape generator.

  • Promising visible-surface reconstruction results. The method achieves consistent improvements in normal-map similarity and angular accuracy across two benchmarks. The qualitative results also demonstrate better preservation of certain geometric structures, particularly for objects with complex surfaces and thin components.

Critical weaknesses
  • Insufficient ablations to isolate the proposed method's contribution. The main comparison uses native Pixal3D, whereas SUPRA-Vox is additionally fine-tuned on the curated HQ100k dataset. Therefore, it is unclear how much of the improvement comes from the proposed Native Voxel Prompts versus additional fine-tuning and training data. Although the authors provide an ablation that removes the surface conditions during inference, this introduces a distribution shift rather than evaluating a baseline trained without these conditions. A controlled comparison with a similarly fine-tuned Pixal3D baseline is necessary to establish the effectiveness of the proposed method.

  • Limited improvements in whole-mesh geometry and unclear advantages over simpler alternatives. While the visible-surface metrics improve consistently, the gains in complete geometry over Pixal3D are relatively small (e.g., Chamfer distance decreases from 0.0411 to 0.0386, and F-score increases from 24.63% to 25.60%). The paper also lacks comparisons with simpler alternatives for incorporating estimated depth or partial geometry. Given the existing literature on geometry-conditioned 3D generation and native latent inpainting, it remains unclear whether the additional complexity of the proposed approach is justified.

Questions for authors
  • Can the authors compare SUPRA-Vox against Pixal3D fine-tuned on the same HQ100k dataset with an identical training budget but without Native Voxel Prompts? This would help isolate the contribution of the proposed conditioning mechanism.

  • How does Native Voxel Prompt conditioning compare with simpler alternatives, such as directly concatenating projected depth features or binary voxel occupancy? What specific advantage does encoding the observed surface through the pretrained VAE provide?

  • Can the authors provide additional evidence of robustness, including uncertainty estimates for the whole-mesh improvements and results across different random seeds? It would also be useful to report whole-mesh geometry metrics on DTC-Geo.

Reviewer_VAL1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SUPRA-Vox presents a surface-guided framework for single-image 3D asset generation. It converts a monocular geometry estimate into native voxel prompts and combines them with a spatial mask to guide sparse-structure generation throughout denoising. The resulting structure is passed to a frozen downstream generator for detailed mesh synthesis. The method improves reference-view surface metrics over the evaluated baselines and improves some whole-mesh metrics over native Pixal3D.

Critical strengths

The paper is well-written, and its innovative aspects are quite easy to grasp.

A clear and relevant objective. The paper targets visible-surface fidelity, addressing missing parts and misplaced geometry in single-image 3D generation.

A practical integration of geometric guidance. Encoding the estimated surface in the generator’s native structure space allows the method to use explicit geometry while retaining the pretrained downstream generator.

Critical weaknesses

The contribution of the prompts is not fully isolated. SUPRA-Vox is fine-tuned on additional 3D data, whereas the main Pixal3D baseline uses its original checkpoint.

The alignment claim is stronger than the whole-mesh evaluation supports. Whole-mesh metrics are computed after ground-truth-assisted Sim(3) alignment. They therefore measure shape quality after alignment and cannot independently establish how well the original output aligns with the input image.

The method can inherit errors from the geometry estimator. Training uses rendered ground-truth depth, while inference uses estimated depth. The paper shows a case where incorrect surface evidence persists in the generated result, but does not systematically measure sensitivity to depth or camera errors.

Questions for authors
  1. Can you compare SUPRA-Vox with fine-tune Pixal3D on the same data and budget without voxel prompts?

  2. Can the authors compare Native Voxel Prompts with simpler conditioning strategies, such as directly encoding the predicted depth map or point cloud, under a comparable training budget?

  3. Can you report results without ground-truth-assisted alignment?

  4. How does performance change with depth noise, missing regions, or camera errors?

  5. Can confidence-aware prompting reduce failures caused by inaccurate surface estimates?

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

Related papers

Discussion (0)

Sign in to comment.