acceptodds
Under review as a conference paper at ICLR 2027

Entro-Bridge: Entropy-Structured Refinement for Feed-Forward 3D Reconstruction

Abstract

Recent advances in single-view 3D reconstruction have been enabled by feed-forward and generative models, allowing efficient learning-based inference from incomplete visual observations. However, these models often overlook the spatially heterogeneous reliability of reconstructed geometry: feed-forward predictions may contain reliable observations, uncertain predictions, and unsupported regions, while generative completion may overwrite reliable structures during unseen-region recovery. We formulate this challenge as a reliability-aware stochastic repair problem, where reliable geometry should be preserved, uncertain predictions should be corrected, and unsupported regions should be completed. To solve this problem, we propose Entro-Bridge, an entropy-structured paired Schrodinger Bridge framework that enables selective geometric repair through uncertainty-controlled stochastic transport. Specifically, we introduce an Entropy-structured Reliability-aware Source Boundary (ERSB), which transforms feed-forward predictions into a source distribution with spatially structured entropy: reliable observations are assigned low uncertainty, uncertain predictions allow greater flexibility for correction, and unsupported regions retain greater freedom for completion. During reverse transport, the Reliability-Guided reverse Repair Dynamics (RGRD) carries this reliability structure into the reverse repair process, guiding the correction trajectory toward uncertain regions while preserving reliable geometry. On the ABO and 3D-FRONT datasets, Entro-Bridge consistently outperforms state-of-the-art methods, reducing Chamfer Distance by approximately 13.0 % and 10.3% relative to SOTA methods, with F-scores of 92.0 % and 81.3 %, respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Single-view 3D Reconstruction, Feed-forward 3D Reconstruction, 3D Shape Completion, Schrödinger Bridge, Reliability-Aware Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_dBD9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Entro-Bridge repairs single-image feed-forward geometry using a frozen VGGT source and a frozen TRELLIS representation and downstream generator. Observation support and confidence control source-latent retention and Gaussian noise injection. A supervised paired bridge trains a repair network, while region-specific objectives and reliability-aware local geometric attention encourage preservation, correction, and completion in different regions. Five reverse steps produce the repaired sparse structure. Object and scene models are trained on Toys4K and HSSD and evaluated on ABO and 3D-FRONT. The main contribution is reliability-aware latent repair built on established paired-bridge machinery.

Critical strengths
  1. The method distinguishes reliable observations, uncertain predictions, and unsupported geometry, rather than treating all partial geometry as equally trustworthy. Using this distinction in both the source endpoint and repair network addresses the important conflict between completing missing structure and rewriting useful observations.
  2. Table 3 directly evaluates the claimed completion-preservation balance through GT-defined unseen-region metrics, visible-region GT error, and drift from shared source anchors. Controlled bridge variants and confidence-perturbation experiments provide additional relevant evidence. This is more informative for the central claim than reporting global reconstruction accuracy alone.
  3. The repair is implemented as a compact module with five reverse steps and 6.958M trainable parameters, while the foundation models remain frozen. The paper distinguishes the 0.606 s repair-stage measurement from the 8.361 s complete-pipeline runtime, making the efficiency claim interpretable. The appendix also carefully separates fixed paired coupling, stochastic training, and deterministic inference, avoiding an unsupported claim of directly solving the unrestricted marginal bridge problem.
Critical weaknesses
  1. The quantitative record is internally inconsistent and important evaluation details are missing. Table 1 reports the ABO result with standard deviations 0.0004 for CD, 0.28 for F-score, and 0.110 for PSNR-N; Appendix D, Table 4 gives the same means with deviations 0.0010, 0.91, and 0.378. Several baseline F-score deviations also change. Different aggregation levels might explain this, but none is specified. Figure 4 plots CD near 0.201, whereas the tables report values near 0.020 and Appendix D states that no multiplicative scaling is applied. Fixed test-set sizes or manifests are not supplied, and the strongest Points-to-3D baseline is reimplemented and trained without sufficient recipe detail to verify equivalence. These issues materially limit confidence in the claimed advantage, without establishing fabrication.
  2. The independent contribution of the entropy/bridge formulation remains incompletely established. The Gaussian entropy derivative describes the chosen noise parameterization, and the fixed-coupling KL decomposition provides a standard conditional-bridge interpretation; neither proves geometric correctness after learning and decoding. Improvements over Vanilla Paired SB are encouraging, but the precise control configurations are not fully specified. Replacing RLGA with global self-attention simultaneously changes locality, geometry, and reliability, while fine-grained results suggest a larger effect from geometry bias than confidence bias. Variance-level endpoint mixing is also relatively close to the full model. The contribution could be valuable algorithmically, but stronger matched controls are needed to distinguish it from general confidence-aware local repair and to identify what the bridge formulation adds.
Questions for authors
  1. Please reconcile Tables 1 and 4, specify the aggregation unit and source of randomness for every reported standard deviation, and explain the CD scale in Figure 4. Can you provide per-run results and the fixed test-set sizes or manifests?
  2. What training data, source construction, alignment, checkpoints, sampling budgets, and validation checks were used for the Points-to-3D reimplementation? How are native canonical-frame outputs transformed into the shared source coordinate frame?
  3. Which ingredients differ in Vanilla Paired SB and w/o ERSB, and does the gain persist against local geometric attention without reliability modulation or a matched directly supervised confidence-aware repair model? Precise existing control specifications or equivalent results would materially affect the novelty assessment.
Reviewer_teMsOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work focuses on single-view feed-forward reconstruction’s uneven reliability: some geometry is genuinely observed, some is observed but wrong, some was never seen. The paper's point is that generative completion treats all of it as one undifferentiated condition and so sometimes overwrites the parts that were correct. It reframes refinement as preserve / correct / complete in one pass. Concretely, a paired Schrödinger bridge in the frozen TRELLIS voxel latent space. Rather than Gaussian noise, the bridge's source endpoint is the source latent modulated by VGGT confidence (ERSB): an observed cell is interpolated between the source latent and noise according to its confidence, while an unsupported cell falls back entirely to the Gaussian prior. Supervision is then split by region — high-confidence cells anchored to the source latent, the rest to GT — and repair is done by a local window attention likewise modulated by confidence (RLGA). In the experiments, only the 6.96M-param repair net is trained on Toys4K/HSSD, and it is evaluated zero-shot on ABO and 3D-FRONT, where it reports roughly a tenth off the CD of the strongest baseline. There are ablations, a confidence-perturbation study, and a set of region-wise metrics that separate completion from preservation.

Critical strengths

1.Decomposing the single-view feed-forward output by reliability is a meaningful and natural problem to pose. Existing refinement work conditions on the partial reconstruction as one undifferentiated input, which implicitly treats its errors as uniform. A single view leaves back-facing regions with no evidence at all, while the visible surface mixes geometry that is essentially correct with geometry corrupted by weak texture, reflections and depth ambiguity. Cutting that into preserve, correct and complete is a reasonable way to handle it.

2.The region-wise “preservation and completion” evaluation is helpful to understand the key claims. In Tab. 3 Entro-Bridge is state of the art on Unseen CD (0.021), Unseen F (87.4), Observed GT Error (0.0070) and Anchor Drift (0.0073) — i.e. it completes more accurately while staying faithful to what was already there, which is exactly the trade-off global CD hides. Anchor Drift is well constructed: built from the shared source and its confidence only, no GT and no model output, so it's genuinely comparable across methods.

Critical weaknesses
  1. My main concern: as constructed, C cannot express "reliable," so the paper's core mechanism — three graded levels of reliability — is problematic; and without it, the method sits very close to existing work. Per App. B.1, the 64³ confidence grid is average-pooled down to 16³ with empty child voxels counted as zero. Each latent cell's C is therefore the mean VGGT confidence over its occupied children, multiplied by the fraction of its 64 children that are occupied. A single-view point cloud is a thin surface, which fills at most about a quarter of a 4×4×4 block. So even where VGGT is fully confident, C stays below roughly 0.25, and how it varies from cell to cell is governed mainly by how much surface happens to cross each cell, with no appreciable influence from VGGT's confidence. The regime the paper's entire design is organised around — high-confidence cells acting as near-deterministic anchors — therefore never arises, in my understanding. Every observed cell's source endpoint is at least ~75% noise, RLGA's receiver gate is ≥0.75 everywhere, and the "low-confidence" correction loss effectively covers the whole observed region. In practice ERSB separates observed from unobserved cells with a surface-density weighting on top, rather than distinguishing reliable / uncertain / unsupported. Moreover, no experiment separates the VGGT-confidence factor from the occupancy factor, so whatever gain C does provide cannot be attributed to reliability. And the confidence-graded three-way source endpoint is the idea that distinguishes this work from others. Take it away, and what remains is close to two existing lines: BridgeShape, which already transports partial shapes to complete ones with a latent Schrödinger bridge and supplies the transport equation used here; and training-free latent completion (LaS-Comp), which already holds observed cells near the input and lets missing cells start from noise inside a frozen latent generator, with nothing trained at all — LaS-Comp doing so in this same TRELLIS space. The second line is neither cited nor compared against.

  2. I also have concerns about the reliability of the reported experiments. Tabs. 1 and 4 report the same experiment with identical means but different standard deviations — and the discrepancy shows up specifically in the proposed method's row, in the direction that favours it.(1)The F-score column. The means agree for all nine methods; the σ differ for all nine. (2)The Ours row. The proposed method's σ differ on all four shared metrics (CD ±0.0004 vs ±0.0010, F-score ±0.28 vs ±0.91, and likewise for the two normal-rendering metrics). I would like the authors to clarify this.

Questions for authors
  1. What is the empirical distribution of C over supported cells on ABO (mean / median / 95th pct), and of n_i? If it's concentrated below ~0.25, how does that square with the C → 1 anchor regime? Would you report a variant with C_i normalized by occupied-child count?

  2. To separate reliability from density: what's the correlation between C_i and whether that source cell is actually within τ of GT? And could you add a Tab. 7 control using the pure density field n_i/64 with the VGGT confidence discarded? If that matches the full model, the interpretation needs revising.

  3. Please position the method against, and compare with, the training-free latent-space completion line, where region-dependent noise scheduling over a frozen latent generator is already established.

  4. Please explain the Tab. 1 vs Tab. 4/5 discrepancy

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

This paper proposes Entro-Bridge, a reliability-aware framework for refining single-view 3D reconstructions. Partial geometry predicted by a frozen VGGT model is encoded into the latent space of a frozen TRELLIS encoder. The proposed Entropy-structured Reliability-aware Source Boundary (ERSB) uses observation support and confidence to control source retention and noise injection. A paired conditional bridge, region-specific supervision, and reliability-aware local geometric attention then jointly preserve reliable geometry, correct uncertain predictions, and complete unsupported regions. The paper reports improvements over the evaluated baselines on ABO and 3D-FRONT, supported by component ablations, confidence perturbation experiments, region-wise evaluation, and efficiency analysis.

Critical strengths
  1. The problem formulation is well motivated and leads to a coherent method. Reliable observations, uncertain predictions, and unsupported regions require different treatment in single-view reconstruction. The paper incorporates this distinction into source construction, supervision, and information propagation. This provides a reasonable inductive bias for balancing geometric preservation and completion.
  2. The evaluation provides relevant evidence for the completion–preservation objective. Table 3 jointly examines unseen-region completion, visible-region accuracy, and Anchor Drift. Importantly, the visibility partition is constructed from ground-truth geometry and the input camera independently of the method’s internal masks. Evaluating visible-region accuracy alongside source preservation helps distinguish accurate reconstruction from merely retaining source errors. The component ablations and confidence perturbation experiments provide additional encouraging evidence.
Critical weaknesses
  1. The central reliability signal conflates prediction confidence with local occupancy density. Appendix B.1 constructs latent confidence by averaging over 4 x 4 x 4 voxel neighborhoods, with empty voxels contributing zero. Consequently, the resulting confidence depends on both the confidence of occupied voxels and their occupancy fraction. For example, a neighborhood containing only one occupied voxel with confidence 1 receives a latent confidence of 1/64, substantially suppressing source retention in Equation 6. This raises the possibility that sparse or thin structures are treated as unreliable because of low occupancy rather than inaccurate predictions. This does not establish that the construction is ineffective, but it leaves the central reliability interpretation insufficiently validated. The confidence-shuffling experiment in Table 7 does not separate these effects. Direct evidence relating pooled confidence to geometric error, together with an occupied-voxel-only averaging control, would help resolve this concern.
  2. Unexplained statistical inconsistencies and missing evaluation details limit confidence in the reported improvements. For the full model on ABO, Table 1 reports an F-score of 92.00+-0.28, whereas Table 4 reports 92.00+-0.91. The corresponding CD standard deviations are 0.0004 and 0.0010. On 3D-FRONT, the F-score standard deviation changes from 0.36 in Table 1 to 0.98 in Table 5. Different aggregation procedures or experimental runs could explain these discrepancies, but no such distinction is documented. In addition, the sizes and selection criteria of the fixed test sets are not clearly specified. The strongest comparison baseline, Points-to-3D, is reimplemented and trained by the authors, but its training configuration and deviations from the published method are insufficiently documented. These details are necessary to assess the stability and reproducibility of the reported gains.
Questions for authors
  1. How does the pooled confidence correlate with actual source geometry error, particularly for sparse or thin structures? How do reconstruction quality and Anchor Drift change when confidence is averaged only over occupied child voxels?
  2. Why do the standard deviations differ between Tables 1, 4, and 5? Which sources of randomness vary across the three runs, and how are the statistics aggregated?
  3. How many instances are included in the ABO and 3D-FRONT test sets, and how were they selected? What training data, training budget, and input point clouds were used for the Points-to-3D reimplementation relative to the proposed method?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.