acceptodds
Under review as a conference paper at ICLR 2027

One Framework, Diverse Information Deficiencies: Towards Unified 3D Gaussian Reconstruction under Imperfect Observations

Abstract

Real-world 3D reconstruction is often limited by diverse imperfect observations, ranging from incomplete scene coverage to degraded image quality. Despite their different manifestations, we argue that they share a common underlying problem: insufficient reliable scene information for faithful reconstruction. We unify them under the notion of information deficiency and introduce UniFill3D, a unified framework for 3D Gaussian reconstruction under imperfect observations. UniFill3D couples scene-grounded reconstruction, which preserves scene-specific geometry and multi-view consistency, with external prior recovery, which recovers missing, degraded, or unreliable visual information. Their target-aligned predictions are reconciled through hierarchical scene–prior fusion, combining dual-latent single-step diffusion with multi-scale prior-guided decoding. The resulting restored views are then fed back to refine the 3D Gaussian representation. We validate UniFill3D on four representative settings spanning incomplete and degraded observations: sparse-view reconstruction, 3D super-resolution, 3D deblurring, and 3D low-light enhancement. UniFill3D achieves state-of-the-art performance across all four tasks, with ablations confirming the complementary value of scene grounding and external priors. Together, these results support viewing seemingly disparate 3D reconstruction problems as different manifestations of a shared information-deficiency challenge that can be addressed through a common reconstruction paradigm.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Reconstruction, 3D Gaussian Splatting, 3D Generation

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

The paper introduces UniFill3D, a unified framework for 3D Gaussian Splatting (3DGS) reconstruction under diverse imperfect observation conditions framed collectively as information deficiency. The approach produces two target-aligned estimates for each camera: a scene-grounded Reconstruction Anchor rendered from the current 3D representation and a Prior Prediction generated by an external restoration or generative model; these are reconciled by a hierarchical fusion module that combines dual-latent joint attention in a single-step diffusion U-Net with multi-scale prior-guided VAE decoding, after which the restored views supervise 3D refinement.

Critical strengths
  • The target-aligned, two-branch abstraction is a clean, reusable interface that decouples deficiency-specific recovery from shared reconciliation.

  • The fusion mechanism combines a dual-latent joint-attention U-Net with single-step diffusion and multi-scale prior-feature injection into the VAE decoder, providing a principled, lightweight alternative to multi-step sampling.

Critical weaknesses
  • The core idea using external priors to refine 3D reconstructions via target-aligned fusions and feeding restored views back is close to several prior pipelines, making the incremental novelty of the proposed dual-latent single-step diffusion + VAE feature injection the main technical delta.

  • No explicit mechanism for reliability estimation or gating of the prior; if Pq hallucinates off-geometry content, the fusion may propagate artifacts into 3D without uncertainty-aware weighting.

Questions for authors

For sparse-view, how do you choose which cameras receive pseudo-observations and how many? Is performance sensitive to the schedule of adding restored views and to the relative weighting of restored vs. real views during 3D refinement?

5 comments
AuthorsOpenReview ↗

Response to Reviewer TRcd (Part 1/2)

We thank the reviewer for recognizing the two-branch interface and fusion mechanism. We provide the pseudo-observation settings and additional tests of supervision weights and refresh schedules below.

Questions

Q1. Pseudo-Observation Cameras, Counts, Introduction Schedule, and Supervision Weights

For sparse-view, how do you choose which cameras receive pseudo-observations and how many? Is performance sensitive to the schedule of adding restored views and to the relative weighting of restored vs. real views during 3D refinement?

Pseudo-observations follow the progressive camera path of DiFiX3D+: cameras advance from near the input views toward the target poses. Each round rerenders the scene, obtains external predictions, and performs fusion at those camera positions. This process uses target camera poses.

Each round contains one pseudo-observation per target camera. The count stays fixed within a scene, while camera positions and coverage change. The main settings are:

ItemSetting
Pseudo-observations per roundOne per target camera; 39–52 per scene across the 12 DL3DV scenes.
3D optimization intervalContinue from the 30K scene state to 60K.
Refresh scheduleOne round every 2K updates, for 15 rounds.
Pseudo-supervision updateEach round replaces the previous pseudo-views; real observations are retained.
Real/pseudo-observation sampling probabilities0.7/0.3.
Real/pseudo-observation loss weights1.0/0.3.

Sampling probability controls how often a supervision source participates; the loss weight controls its strength in that update.

Sensitivity to the Introduction Schedule and Relative Supervision Weight

We completed two sets of controls on three preselected DL3DV scenes with 148 target views. All other settings match the paper's progressive refinement configuration. All metrics are measured after final query-view fusion, matching the paper's sparse-view evaluation.

The weight test uses a uniform schedule and pseudo-view loss weights of 0.15, 0.30, and 0.60. The real-view weight remains 1.0, and the pseudo-view sampling probability remains 0.3. The schedule test fixes the original weights and 15 refreshes, comparing uniform spacing with schedules that refresh more frequently early or late in optimization. More frequent early refreshes use seven 1K intervals followed by seven 3K intervals; more frequent late refreshes reverse this order.

ConfigurationFinal fusion PSNR ↑Final fusion SSIM ↑Final fusion LPIPS ↓
Weight 0.15, uniform15.6830.58780.4928
Original: weight 0.30, uniform15.6800.58770.4927
Weight 0.60, uniform15.6570.58760.4930
Weight 0.30, more frequent early refreshes15.6710.58740.4924
Weight 0.30, more frequent late refreshes15.7140.58850.4933

Relative to the original configuration, the absolute change in mean PSNR for final fusion is at most 0.035 dB. This small-scale test suggests that the method does not depend on a precise weight or uniform refresh intervals.

AuthorsOpenReview ↗

Response to Reviewer TRcd (Part 2/2)

Weaknesses

W1. Relationship to Existing Pipelines and Technical Contributions

UniFill3D follows the restore-then-refine paradigm and uses the DiFiX3D+ backbone and checkpoint initialization. Our focus is fusing the current reconstruction and external predictions within this procedure.

We construct a Reconstruction Anchor and a Prior Prediction for the same target camera, then combine them through dual-latent interaction and multi-scale feature injection during decoding. For sparse views, a video generation model supplies additional information at unobserved views, and the current Gaussian rendering provides the corresponding scene estimate. Fused views then supervise 3D refinement. The independently trained ablations in the submission show that full fusion outperforms either branch alone on both sparse-view reconstruction and super-resolution.

This target-aligned fusion interface, architecture, and training objective can be reused across tasks and are not tied to specific internal reconstruction or external prior models. A reconstruction model and a generative or restoration model can replace the respective branches if their predictions correspond to the same target camera and image plane. Replacement requires retraining the fusion weights, while the shared fusion architecture and training objective remain unchanged. Different model combinations can therefore fuse their predictions through the same interface and use the restored views for 3D optimization.

W2. Prior Reliability and Hallucination Propagation

During fusion, the updated Reconstruction Anchor latent and prior features jointly participate in decoding. During sparse-view refinement, real observations continue to constrain the Gaussian representation jointly optimized across views. We rely on these two stages to constrain prior use; joint 3D optimization is intended to suppress residual inconsistencies among restored views.

The current method has no separate confidence estimate, geometry-aware gate, or conflict-based view rejection. Adding reliability prediction and geometric checks is a possible extension for reducing prior-error propagation.

Please let us know if any point needs further clarification or if you have additional questions. We will respond promptly. Thank you again for your time and thoughtful feedback.

AuthorsOpenReview ↗

Follow-Up on Rebuttal Discussion

Dear Reviewer TRcd,

Thank you again for your time and valuable feedback. We have provided point-by-point responses, additional experimental results, and clarifications in the revised manuscript to address your concerns.

As the discussion period draws to a close, we would appreciate hearing whether any concerns remain unresolved or require further clarification. We would be happy to provide additional explanations during the remaining discussion period.

Once you have had an opportunity to consider our responses and additional results, we would also appreciate your updated assessment.

Thank you again for your time and consideration.

Best regards, Paper26205 Authors

Reviewer_V7vLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces UniFill3D, a framework for 3D Gaussian reconstruction from incomplete or degraded observations. For a target camera, a scene-grounded reconstruction is rendered from a current 3D representation and an external prior model produces a complementary prediction. Both predictions are aligned to the same target camera and image plane. A hierarchical scene–prior fusion module combines them through joint interaction between two VAE latent streams, a single diffusion U-Net update, and multi-scale features from the prior prediction injected into the VAE decoder. The restored target views are then used as supervision for a subsequent optimization of the underlying Gaussian representation using the native 3D reconstruction objective. The paper instantiates this pipeline for three-view sparse reconstruction, four-times 3D super-resolution, 3D deblurring, and low-light enhancement. A separate fusion checkpoint is trained for each task, while the architecture and training objective are kept the same.

Critical strengths
  1. The target-aligned scene–prior interface is well motivated. Two predictions for the same camera provide a sensible way to combine scene fidelity with recovered information.

  2. The fusion architecture has a coherent two-level design. Joint latent interaction enables information exchange, while multi-scale prior features provide spatial detail during decoding. The updated anchor latent keeps the output tied to the scene branch.

  3. The restoration-to-3D interface is simple. Fused images retain their camera poses and can be added through the native 3D objective without changing the Gaussian representation.

Critical weaknesses
  1. The four demonstrations share a fusion architecture and objective, but use task-specific reconstruction and prior models and separately trained fusion checkpoints. The evidence therefore supports reuse of a common fusion recipe across these settings, while the extent to which the framework reduces task-specific design is more limited than its broad framing suggests.

  2. The fusion stage has no additional explicit mechanism for multi-view consistency. Native 3D refinement provides an indirect constraint, but the paper does not describe confidence weighting or cross-view checking before fused views become supervision. Prior-induced inconsistencies must therefore be handled by the native objective alone.

  3. The method does not explicitly control the relative reliability of scene and prior information. The prior affects the output through joint attention and multi-scale injection, but no uncertainty estimate, spatial confidence map, or geometry-aware gate is described. The scene anchor alone does not prevent prior hallucinations.

  4. Restoration-to-3D refinement can propagate fusion errors into the Gaussian representation. The paper does not describe robust weighting or rejection when fused outputs disagree with observed views or one another.

  5. The fixed single-step diffusion configuration is under-justified across the four deficiencies. All tasks use the same timestep and the instruction “remove degradation,” despite different inverse problems. The paper does not explain why these settings are appropriate or how sensitive the fusion is to them.

Questions for authors
  1. Does refinement give restored views the same weight as original observations, or does it downweight or reject fused targets when they conflict with the scene anchor or other target views?

  2. Why are the same diffusion timestep and instruction used for all four deficiencies, and what in the task-specific conditioning compensates for the different inverse problems?

7 comments
AuthorsOpenReview ↗

Response to Reviewer V7vL (Part 1/2)

We thank the reviewer for recognizing the target-aligned interface and hierarchical fusion design. Below we explain the supervision settings, rationale for the shared configuration, and additional experiments.

Questions

Q1. Supervision Weights for Restored Views and Conflict Handling

Does refinement give restored views the same weight as original observations, or does it downweight or reject fused targets when they conflict with the scene anchor or other target views?

In sparse-view optimization, real observations and restored views are sampled with probabilities 0.7 and 0.3, with loss weights 1.0 and 0.3, respectively. Real observations therefore provide the primary supervision, while restored views participate less frequently and with a lower weight.

The current method does not dynamically downweight or reject supervision based on view conflicts. We rely on continued supervision from real observations and joint optimization of the shared 3DGS to constrain residual inconsistencies: restorations from different cameras supervise the same 3D representation, while real observations continue to provide scene constraints.

For super-resolution, deblurring, and low-light enhancement, the observations themselves are degraded. We therefore generate fused restorations at the observed cameras and use them as fixed supervision to continue optimizing the existing 3D representation. We will describe these two supervision settings separately in the revision.

Q2. Rationale and Sensitivity of the Shared Timestep and Instruction

Why are the same diffusion timestep and instruction used for all four deficiencies, and what in the task-specific conditioning compensates for the different inverse problems?

We thank the reviewer for the question. The shared timestep and instruction reflect the role of the diffusion module in our framework. Before fusion, the reconstruction and external prior models for each task have already produced a Reconstruction Anchor and a Prior Prediction at the same target camera. The fusion module addresses a common problem: combining these two predictions to restore the target view.

Task information enters fusion primarily through the image conditions. The Reconstruction Anchor provides the current scene estimate, while the Prior Prediction supplies task-relevant restored or generated content, such as high-resolution details for super-resolution or novel-view predictions from a video generation model for sparse-view reconstruction. Separately trained fusion weights further adapt to the error characteristics and complementary information of the two predictions in each task.

We therefore use the generic instruction “remove degradation” to express the shared view-restoration goal and retain t=199 as the common single-step fusion configuration.

Following your suggestion, we added inference tests with other parameters fixed. The timestep test uses 120 input pairs from 12 sparse-view scenes. Scene states after 3D optimization and both predictions remain fixed; only the timestep of final query-view fusion changes.

TimestepPSNR ↑ΔPSNR relative to t=199 (dB)
9916.595+0.037
14916.576+0.019
19916.5570
24916.319−0.238

The PSNR at t=199 is approximately 0.037 dB below the highest tested value. Increasing the timestep further reduces performance: at t=299, PSNR is 16.028, a decrease of 0.529 dB from the original configuration.

The instruction test uses 120 input pairs per task at t=199, comparing “remove degradation,” “restore the image,” a task-specific instruction, and an empty instruction.

TaskMaximum absolute ΔPSNR across alternative instructions (dB)
Sparse-view reconstruction0.0165
Super-resolution0.0056
Deblurring0.0287
Low-light enhancement0.0557

Across all four tasks, the maximum PSNR change from changing the instruction is less than 0.056 dB, supporting continued use of a generic restoration instruction.

AuthorsOpenReview ↗

Response to Reviewer V7vL (Part 2/2)

Weaknesses

W1. Scope of the Unified Framework and Task-Specific Design

The target-camera interface, fusion architecture, training objective, and subsequent 3D optimization procedure are shared. Each task uses its corresponding reconstruction model, external prior, and independently trained fusion weights. We will use “a unified reconstruction framework with task-specific instantiations” to distinguish the shared and task-specific parts.

The framework reuses how the two predictions enter fusion and return to 3D optimization, reducing repeated design of the fusion interface and architecture. The internal model supplies a prediction from the scene representation, while the external model supplies a generation or restoration prediction at the same target camera and image plane. These predictions need not come from specific models. Models satisfying the interface requirements can replace the corresponding branches and use the same fusion and 3D optimization procedure after retraining the fusion weights. Each task can therefore select models suited to its degradation while retaining the shared fusion architecture and training objective.

W2, W3. Reliability and Cross-View Checks at the Fusion Stage

During fusion, the updated Reconstruction Anchor latent and multi-scale prior features jointly participate in decoding, keeping scene estimates involved in restoration. The independently trained single-branch ablations in the submission show that full fusion outperforms either branch alone on sparse-view reconstruction and super-resolution, supporting their complementarity.

There is no separate uncertainty estimate, confidence map, or geometry-aware gate, and no cross-view filtering before fused views become supervision. Residual inconsistencies are constrained through the joint 3D optimization described in Q1. Reliability prediction and geometric checks could further control prior errors.

Please let us know if any point needs further clarification or if you have additional questions. We will respond promptly. Thank you again for your time and thoughtful feedback.

The response clarified the supervision weights and provided a useful sensitivity analysis. It also defined the framework more precisely as shared fusion and optimization components while using task-specific models and checkpoints. However, the fusion stage did not dynamically assess prior reliability or filter conflicting restoration views before using them as supervision. Is it possible that the restoration views contain conflicting or unreliable prior details? If so, how does the joint 3D optimization handle these details?

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

This paper proposes "information deficiency" as a unifying perspective for 3D Gaussian reconstruction under imperfect observations, including sparse views, low resolution, blur, and low illumination. The presented UniFill3D framework combines two complementary pathways: a scene-grounded reconstruction branch that preserves scene geometry and multi-view consistency, and an external prior recovery branch that restores missing or degraded visual information. A hierarchical scene–prior fusion module — integrating dual-latent single-step diffusion and multi-scale prior-guided decoding — reconciles the two predictions, and the restored views are then fed back to refine the underlying 3D Gaussian representation. The method achieves state-of-the-art results across four heterogeneous reconstruction tasks, with ablations confirming the complementary value of the two branches.

Critical strengths

The information deficiency perspective offers a useful unifying lens for traditionally isolated imperfect-reconstruction tasks, providing conceptual insight for the community.

Critical weaknesses
  1. The "unified framework" claim is overstated. Each task uses a separately trained fusion checkpoint paired with task-specific anchor and prior models. Only the architectural template is shared, rather than a single end-to-end multi-task model, making the unification largely methodological rather than functional.
  2. Technical novelty is incremental. The overall "restore-then-refine" paradigm follows prior work (e.g., Difix3D+), and the fusion module is initialized from the DiFix3D+ checkpoint. Improvements are primarily engineering refinements rather than fundamental methodological advances.
  3. Generality is insufficiently validated. Only isolated degradation types are tested; compound deficiencies common in real-world scenarios are not explored. The ease of extending the framework to new degradation types is also not demonstrated.
  4. Computational cost (training/inference latency, GPU memory) relative to baselines is not reported, and some recent competitive methods are absent from comparisons.
Questions for authors
  1. Beyond extending to more task settings, what are the core technical innovations of the fusion mechanism compared to DiFix3D+?
  2. Could you add fine-grained ablations to quantify the contribution of each key component: single-step diffusion, dual-latent joint attention, and multi-scale prior-guided decoding?
  3. What is the additional computational overhead (training/inference time, GPU memory) compared to vanilla 3DGS and task-specific baselines?
5 comments
AuthorsOpenReview ↗

Response to Reviewer YezJ (Part 1/3)

We thank the reviewer for recognizing the information-deficiency perspective and for the suggestions on technical contributions and experimental analysis.

Questions

Q1. Contributions of the Fusion Mechanism Compared with DiFiX3D+

Beyond extending to more task settings, what are the core technical innovations of the fusion mechanism compared to DiFix3D+?

UniFill3D shares the restore-then-refine paradigm with DiFiX3D+ and uses its checkpoint for initialization. The key difference is that DiFiX3D+ uses real reference observations to repair reconstruction renderings, whereas UniFill3D fuses predictions from the current reconstruction and an external generative model for the same target view.

The two predictions supply different information: the Reconstruction Anchor retains scene structure supported by actual observations, while the Prior Prediction supplies content missing or insufficiently recovered in the reconstruction. For sparse-view reconstruction, for example, a video generation model provides additional information at unobserved views. We align its predictions with Gaussian renderings at the same cameras, allowing the fusion module to compare scene estimates and generated content at corresponding locations.

To combine these sources, we use Hierarchical Scene–Prior Fusion. Dual-latent interaction first exchanges information between the two streams in the diffusion U-Net. Multi-scale features from the Prior Prediction are then injected into the VAE decoder to supply spatial detail for the updated Reconstruction Anchor latent. The fused views subsequently supervise 3D refinement. This design explicitly introduces an external prediction and uses both sources during latent restoration and decoding.

Different tasks can use the same interface and fusion procedure by replacing the reconstruction and prior models with models serving the corresponding roles and training the fusion weights. Our contributions thus include constructing and aligning the two predictions, hierarchical fusion, and reuse of this design across different information deficiencies.

Q2. Roles of the Three Fusion Components

Could you add fine-grained ablations to quantify the contribution of each key component: single-step diffusion, dual-latent joint attention, and multi-scale prior-guided decoding?

We added three inference ablations on the same 14 paired examples, keeping the trained weights fixed. We separately disable single-step latent restoration, dual-latent joint attention, and the feature injection used in multi-scale prior-guided decoding.

The first bypasses the U-Net and scheduler, feeding the original Reconstruction Anchor latent to the decoder while retaining multi-scale prior-feature injection. The second disables only joint attention between the two latent streams, retaining within-stream self-attention and text cross-attention. The third retains single-step latent restoration but disables all four levels of prior-feature injection into the decoder.

ConfigurationPSNR ↑ΔPSNR
Full model15.858—
Bypass single-step latent restoration14.620−1.238
Disable joint attention15.501−0.357
Disable multi-scale prior-feature injection14.361−1.497

Metrics are averaged equally over images; ΔPSNR is relative to the full model. Disabling any pathway reduces PSNR.

AuthorsOpenReview ↗

Response to Reviewer YezJ (Part 2/3)

Q3. Additional Computational Cost Relative to Baselines

What is the additional computational overhead (training/inference time, GPU memory) compared to vanilla 3DGS and task-specific baselines?

UniFill3D adds fusion training, external prediction and fusion, and subsequent 3D optimization.

Across the four tasks, 10K fusion-training updates take approximately 8 hours on four A40 GPUs, with an average peak allocated memory of approximately 34 GiB per GPU. Upstream models use public pretrained weights. Fusing one pair of 1024×576 images takes approximately 1 second and 9 GiB of memory.

We further collected nine cost comparisons across the four tasks. All configurations follow the paper's evaluation settings. Fusion/repair time measures a complete model forward pass that produces one restored image from a pair of inputs. Refinement time covers continued 3D optimization using restored images. Novel-view inference produces the final target view at a given target camera after scene optimization.

TaskMethodFusion / repair (s/pair, approx.)3D refinement (min/scene, approx.)Novel-view inference (s/view, approx.)Fusion / repair module peak memory (GiB)
Sparse-view3DGS——0.0034—
Sparse-viewDiFiX3D+1.3822.51.408.2
Sparse-viewUniFill3D0.7819.111.418.6
3D super-resolutionSRGS——0.0149—
3D super-resolutionUniFill3D0.7815.20.00598.6
3D deblurringDeblurring-3DGS——0.0038—
3D deblurringUniFill3D0.782.20.00418.6
3D low-light enhancementLuminance-GS——0.0030—
3D low-light enhancementUniFill3D0.783.40.00228.6

UniFill3D's sparse-view refinement uses 15 progressive rounds, each invoking the external model before fusion and 3D optimization. It uses two A40 GPUs, with one independently running the external model online. The other tasks use one A40.

Sparse-view final prediction includes Gaussian rendering, SEVA target-view generation, and fusion, adding computation compared with direct rendering by 3DGS or repair using existing real reference images in DiFiX3D+. In this measurement, SEVA batch generation amortized over 45 target views takes approximately 10.6 seconds per view, accounting for approximately 93% of the total inference time of 11.41 seconds per view; fusion takes approximately 0.8 seconds.

Super-resolution, deblurring, and low-light enhancement render novel views directly after scene processing, without further external-model or fusion calls, keeping novel-view inference in the millisecond range.


AuthorsOpenReview ↗

Response to Reviewer YezJ (Part 3/3)

Weaknesses

W1. Scope of the Unified Framework

We thank the reviewer for pointing out the ambiguity in “unified.” We agree that the paper uses task-specific reconstruction models, prior models, and separately trained fusion weights. We will describe UniFill3D more precisely as “a unified reconstruction framework with task-specific instantiations” and clarify this scope in the abstract, introduction, and method sections.

UniFill3D is unified at the level of the reconstruction procedure. Across all four information deficiencies, we align predictions from the scene reconstruction and external prior branches to the same target camera and image plane. We use the same hierarchical fusion architecture and training objective to combine these predictions, then use the restored views to supervise 3D optimization. Dual-latent interaction, multi-scale prior injection, and restoration-guided 3D refinement can therefore be reused across tasks. Task-specific models provide the Reconstruction Anchor and Prior Prediction, while the fusion module learns to combine these two sources through training for each task.

This interface is not tied to specific reconstruction or prior models. Models that serve the corresponding reconstruction, generation, or restoration role and satisfy the target-alignment requirements can replace the respective branch. Replacement requires retraining the fusion weights, while the shared fusion architecture and training objective remain unchanged. The model combinations in the paper are task-specific instantiations, and the four-task experiments support reuse of this common procedure.

W3. Compound Degradations and Extension of the Framework

We added an experiment combining sparse-view reconstruction with 4× super-resolution. Each scene provides only three low-resolution observations at 256×144, with evaluation at novel views at 1024×576. This tests both view completion and detail recovery.

The scene branch uses 3DGS reconstructed from the low-resolution observations. The external branch first super-resolves the inputs with SwinIR and then predicts target views with SEVA. We train fusion weights for this combination while retaining the fusion architecture, training objective, and progressive 3D refinement procedure.

Comparisons include the two predictions alone, SwinIR→FSGS, an adapted FSGS+SRGS combination, and SplatSuRe (SwinIR). All methods use the same three source images, SfM initialization points visible from the source views, and the same 12 DL3DV scenes with 526 target views. GT is used only for evaluation. PSNR, SSIM, and LPIPS are averaged equally across scenes; FID is computed over all target images.

Method / outputPSNR ↑SSIM ↑LPIPS ↓FID ↓
Internal model (3DGS)11.930.3920.667281.66
External model (SwinIR+SEVA)14.900.5390.51750.48
SwinIR→FSGS14.730.5660.614173.83
FSGS+SRGS13.430.4890.628213.28
SplatSuRe (SwinIR)13.580.5070.607191.99
UniFill3D16.160.5670.43945.27

Our method outperforms the other compound-task comparisons on all metrics. SplatSuRe uses SwinIR as its super-resolution model.

The four tasks in the submission cover most mainstream information-deficiency and degradation settings studied in 3D reconstruction. The compound experiment further demonstrates the applicability of the shared procedure. For future degradations, we are confident that the framework can be extended by choosing scene and restoration models for the corresponding roles and training the fusion weights.

W4. Comparisons with Recent Methods

We thank the reviewer for the suggestion. The deblurring experiments in the submission already include PRISM3D (ECCV 2026), and the additional compound-degradation experiment includes a comparison with SplatSuRe (CVPR 2026). We are also evaluating further recent methods under the same input and evaluation settings to broaden the comparisons.

Please let us know if any point needs further clarification or if you have additional questions. We will respond promptly. Thank you again for your time and thoughtful feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Follow-Up on the Rebuttal Discussion

Dear Area Chair,

Thank you for your time in overseeing the review process. We have provided our responses and additional results and remain happy to clarify any outstanding questions.

As we have not yet received follow-up feedback from the reviewers, we are unsure whether any concerns remain. With the discussion period drawing to a close, we would be grateful for any guidance you may have and would welcome the opportunity to provide further clarification if helpful.

Thank you again for your time and consideration.

Best regards, Paper26205 Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.