acceptodds
Under review as a conference paper at ICLR 2027

HashMod: Plug-and-Play Patch-Wise Hash-Guided Refinement for Feed-Forward 3DGS

Abstract

Feed-forward 3D Gaussian Splatting (3DGS) has emerged as an important approach to efficient and generalizable 3D scene reconstruction, typically decoding image features into Gaussian primitives in a single forward pass. However, reconstruction quality may degrade on unseen scenes with different geometric structures, texture patterns. Fine-tuning only the Gaussian head offers a low-cost adaptation strategy, but its ability to correct local geometric and appearance errors can be limited by insufficient emphasis on relevant backbone features.To address this limitation, we propose HashMod, a lightweight modulation module that uses patch-wise binary query–key matching to adaptively select both the number and identity of backbone feature levels contributing to each local correction. HashMod aggregates the selected feature-derived offsets to multiplicatively refine the initial Gaussian predictions. Across three public benchmarks and cross-dataset settings, HashMod achieves competitive or improved reconstruction quality across AnySplat, YoNoSplat, VGGT-, and Depth Anything 3, while introducing only approximately 8M additional parameters and reducing training time by approximately 41–42%. Furthermore, we conduct fine-tuning experiments on two self-collected crop datasets, each comprising 1,000 videos. The adapted feed-forward 3DGS models improve reconstruction quality on real-world crop scenes, demonstrating the practical applicability of HashMod beyond standard benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Feed-forward 3D Reconstruction, Novel View Synthesis, Multi-view Reconstruction, Fine-tuning

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_z8EyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents HashMod, a lightweight plug-and-play refinement module for feed-forward 3DGS. It uses patch-wise binary hash matching to adaptively select relevant backbone feature levels and aggregates the predicted offsets from these levels to refine the original Gaussian parameters. Experimental results show that HashMod generally improves reconstruction quality and achieves comparable performance with fewer fine-tuning iterations.

Critical strengths

The motivation is reasonable, as different local regions may benefit from different levels of backbone features for Gaussian refinement. The proposed method is also simple, lightweight, and easy to follow, making it practical to integrate into existing feed-forward 3DGS models.

Critical weaknesses
  1. The motivation for using binary hashing is not very clear. The idea of selecting different feature levels for different regions makes sense, but why does this selection need to be based on binary hashing? Why not simply use cosine similarity or other conventional similarity measures? Since there are only a few feature levels to choose from, it is also unclear what advantage binary hashing offers. The paper does not discuss this design choice or provide comparisons with these simpler alternatives, making it difficult to see why hashing is necessary here.
  2. It is unclear whether the improvements actually come from hash-guided selection. HashMod adds extra refinement heads, which also introduce more parameters. Since the performance gains are quite small, I wonder if simply adding these refinement heads would achieve similar results even without hashing. The paper does not provide a fair comparison to clarify this.
Questions for authors
  1. Could the authors compare binary hashing with simpler alternatives, such as cosine similarity or other conventional similarity measures, to show why binary hashing is necessary?
  2. Could the authors provide a baseline that simply adds the same refinement heads without hash-guided selection? This would help clarify whether the improvements come from hashing or just the extra refinement heads.
  3. In Fig. 3, the zoomed-in details are only shown for the models without HashMod, rather than comparing the same model with and without HashMod. Could the authors provide side-by-side zoomed-in comparisons? The current figure makes it difficult to see what HashMod actually improves.
3 comments

Thank you for your constructive review and for recognizing the motivation and practical design of HashMod. To address your concerns, we have added cosine-matching and adaptive-gating comparisons, a baseline with the same refinement heads but no selection, and paired zoomed-in visualizations. We hope these additions clarify the contribution of the proposed design and would appreciate your reconsideration of the paper based on the updated evidence.

Critical Weaknesses:

W1 Motivation for binary matching

Thank you for this comment. We have added comparisons with continuous cosine matching and parameter-matched adaptive gating to evaluate this design choice directly. The observed benefit lies in lower training reconstruction loss and modest differences in final quality, rather than an advantage in searching over many candidate levels. We will clarify the empirical motivation for binary matching. The settings and results are detailed in our response to Question 1 below.

W2 Contribution of the additional heads

Thank you for this suggestion. We have added a baseline retaining the same four refinement heads while summing all corrections without selection, with n = 4 at every patch. Its final reconstruction quality is close to HashMod, while HashMod has lower training reconstruction loss and improvements at some intermediate checkpoints. The detailed comparison is provided in our response to Question 2 below.

Q1 Comparison with cosine and gating

Thank you for this suggestion. We have added both cosine matching and parameter-matched adaptive gating baselines. The cosine variant removes query/key binarization and computes continuous cosine similarity, keeping the remaining architecture unchanged. The gating baseline retains the same four refinement heads and routing-network parameter count, replacing binary matching and discrete selection with continuous matching and independent sigmoid gates. HashMod has lower training reconstruction loss at every recorded epoch. Under the same training budget, the epoch-10 test results are shown below.

MethodPSNR ↑SSIM ↑LPIPS ↓
Cosine matching21.4850.7320.301
Adaptive gating21.5260.7330.301
HashMod21.5350.7330.300

The cosine comparison examines binary versus continuous matching, while the gating comparison evaluates the routing design under matched parameter capacity. HashMod improves all three final metrics over cosine matching and achieves comparable quality to gating, with slightly better PSNR and LPIPS. These results provide empirical support for the chosen matching design, with the clearest observed difference in training reconstruction loss. Implementation details and final results are provided in Appendix A.4 (Comparison with Alternative Routing Mechanisms), Table A2. Full curves are provided in Appendix A.3 (Training Convergence and Equal-Iteration Comparisons), Fig. A2.

Q2 Same heads without selection

Thank you for this suggestion. We have added HashMod w/o Selection, which retains the same four refinement heads but sums all their corrections without hash-guided selection, fixing n = 4 at every patch. The multiplicative refinement formulation and training settings remain unchanged. This controls for the additional refinement heads when assessing the contribution of selection. HashMod achieves lower training reconstruction loss at every recorded epoch. The test curves also show improvements at some intermediate checkpoints. For example, at epoch 8, HashMod achieves 21.516 dB PSNR, 0.732 SSIM, and 0.301 LPIPS, compared with 21.460 dB, 0.731, and 0.304 without selection. At epoch 10, the two variants reach comparable reconstruction quality.

MethodPSNR ↑SSIM ↑LPIPS ↓
HashMod w/o Selection21.5320.7330.300
HashMod21.5350.7330.300

These results indicate that selection's observed benefit is more evident in training optimization than in the final reconstruction scores. The same heads can reach similar final quality when all levels are aggregated, while adaptive selection lowers training reconstruction loss and provides improvements at some intermediate checkpoints. The baseline and matched-epoch results are detailed in Appendix A.5 (Same Refinement Heads without Selection), Table A3. Complete training and test curves are provided in Appendix A.3, Fig. A2.

Q3 Paired detail comparisons

Thank you for this suggestion. We have updated Fig. 3 with zoomed-in comparisons for each model with and without HashMod.

Reviewer_z8EyOpenReview ↗

Response to rebuttal

Thank you for the detailed response and the additional experiments and visualizations. The lower training loss and improvements at intermediate checkpoints are encouraging.

As the final results are close to those of adaptive gating and the no-selection baseline, it would be helpful to further clarify the practical benefits of binary hash-guided selection, for example in terms of training efficiency or computational cost, and reflect this nuance in the paper.

Thank you for this helpful follow-up. We have clarified the practical benefits in Sec. 4.1, under “Comparison of parameters and training efficiency.” Table 3 reports approximately 41–42% less total fine-tuning time relative to the original baselines, achieved through fewer training iterations. Appendices A.3–A.5 provide the convergence curves and controlled comparisons with adaptive gating and the no-selection baseline. These sections emphasize the observed benefits in training optimization while noting that the final reconstruction results are comparable, reflecting the nuance you highlighted.

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

Adds a 1.5–8.7M-parameter module to feed-forward 3DGS. Per patch, 8-bit binary keys are built per backbone level and matched against a binary query from image and UV features; an MLP reads the four similarity scores plus RGB and UV and predicts how many levels n to keep, and the mask takes the top-n. Each level has its own head predicting an offset over centre, rotation, scale and colour; selected offsets are summed and applied multiplicatively. Backbone and head are jointly fine-tuned with the module. Tested on AnySplat, YoNoSplat, VGGT-Ω and DA3, plus two self-collected crop datasets.

Critical strengths
  • Architecture-agnostic, and genuinely tested. Four models with different head designs and different usable level counts, under 0.74% added parameters each, backbone frozen. Cross-dataset transfer holds too: trained only on DL3DV, it gives YoNoSplat +0.78 dB on Tanks and Temples.
  • The largest gains are in the application nobody else has. On the self-collected millet and maize sets DA3 gains 0.93 and 0.40 dB, more than anywhere in the benchmark tables.
  • The routing is inspectable. Fig. 5 shows level preference and modulation strength per region.
Critical weaknesses
  • Table 4 undercuts the contribution. High-level 21.41, low-level 21.44, FPN fusion 21.47, HashMod 21.53 dB — and LPIPS is slightly worse than fusion (0.300 vs 0.299). The matching, binarization, learned-n selection and per-level heads together buy 0.06 dB over plain fusion. Neither the STE binarization nor the distribution of predicted n is ablated; no histogram rules out the selector converging to n = 4 nearly everywhere, in which case this is fusion with per-level heads and nothing is being selected.
  • The 41–42% saving is a change of schedule. Table 3: AnySplat 24.4 h at 100K versus 14.5 h at 50K, i.e. 0.29 against 0.24 h per 1K iterations, about 19% slower per step. No inference latency is reported anywhere, despite ~8M parameters added to the forward path.
  • "Plug-and-play" costs the base model 0.6–0.8 dB. The jointly trained checkpoint with modulation disabled gives 20.69 / 20.87 / 20.86 dB at 16 / 24 / 32 views against 21.33 / 21.62 / 21.66 for original DA3. Part of the reported gain is therefore recovery of what joint training broke.
  • The motivating claim is never tested locally. Nothing links selected levels to region type quantitatively.
Questions for authors
  1. What is the empirical distribution of predicted n? If n rarely takes 1–3, what does selection contribute?
  2. Why is the modulation-disabled path 0.6–0.8 dB below original DA3, and what does head-only fine-tuning give?
  3. How do Table 4's 0.06 dB and Table 1's gains behave across seeds?
1 comment

Thank you for your detailed review. We clarify the distinction between head-only fine-tuning and modulation ablations, as well as the scope of training-time savings. We also add selection statistics, routing comparisons, and matched-iteration evaluations. We hope this evidence and the clarifications below support reconsideration of your rating.

Critical Weaknesses:

W1 Contribution and selection ablations

Thank you for this observation. We agree that final gains over FPN fusion are modest and vary across metrics. We added a same-head baseline without selection, continuous cosine matching without binarization, and parameter-matched adaptive gating to assess selection and matching beyond the extra heads.

HashMod achieves lower training reconstruction loss than all three baselines at every recorded epoch. Compared with using all four levels, it improves intermediate-checkpoint reconstruction quality while maintaining comparable final quality. The predicted n distribution confirms active use of feature-level subsets. Selection statistics and the same-head comparison are discussed in Q1, with full curves in Appendix A.3, Fig. A2 and routing comparisons in Appendix A.4, Table A2.

W2 Training time and inference overhead

Thank you for highlighting this distinction. We agree with your inference-time calculation. HashMod’s efficiency advantage lies in reaching competitive reconstruction quality with fewer iterations, reducing total training time as shown in Table 3. The added matched-iteration evaluations further support this advantage, showing consistently lower training reconstruction loss and earlier attainment of competitive quality (Appendix A.3, Fig. A2). We will clarify the scope of the training-efficiency claim in the revision.

W3 Joint training and plug and play integration

Thank you for highlighting this ambiguity. Table 1 measures the net gain over independent head-only fine-tuning, while Table 5 evaluates the effect of modulation within the same jointly trained checkpoint. See Q2 for details. By 'plug-and-play,' we mean compatibility with the existing prediction architecture while retaining its original pathway. We have clarified this terminology in the Introduction and the comparison settings in the captions of Tables 1 and 5.

W4 Region types and selected levels

Thank you for pointing out the need for quantitative local analysis. Fig. 5 illustrates spatial variation in feature-level selection, while the selection statistics (Appendix A.6, Fig. A3 and Table A4) and same-head ablation (Appendix A.5, Table A3) further examine how selection is used and what it contributes. The statistics show that the model actively selects feature-level subsets, and the ablation supports its contribution to training optimization. These findings support adaptive selection, although they do not establish a quantitative correspondence between specific region types and selected levels. The revised discussion presents region-dependent feature requirements as design motivation, while Fig. 5 provides qualitative visualization.

Q1 Distribution of selected level counts

Thank you for this question. We report the distribution of n over successive 5K-step training windows. The proportion of n = 4 decreases from 23.786% in the first 5K steps to 13.676% between 5K and 10K steps, and remains zero in all recorded windows after 10K steps. Patches subsequently select either two or three levels. In the final 95K–100K window, 16.840% select n = 2 and 83.160% select n = 3, giving an average of 2.832 selected levels. Selection of n = 1 is negligible and occurs only early in training.

These observations confirm that selection does not collapse to all-level aggregation. We also train a baseline with the same four refinement heads that sums all corrections without selection, fixing n = 4 at every patch. HashMod achieves lower training reconstruction loss at every recorded epoch with comparable final quality, supporting its benefit in training optimization. See Appendix A.6, Fig. A3 and Table A4 for distributions; Appendix A.5, Table A3 for the same-head baseline; and Appendix A.3, Fig. A2 for full curves.

Q2 Modulation disabled and head only results

Thank you for highlighting this ambiguity. Tables 1 and 5 use different baselines. Table 1 compares HashMod with independently trained head-only fine-tuning under the same 100K-iteration budget, showing net PSNR gains of 0.200 / 0.240 / 0.240 dB for 16 / 24 / 32 views.

Table 5 instead enables or disables modulation within a single jointly trained checkpoint. With the backbone frozen, the Gaussian head and HashMod are jointly optimized through the final modulated output, allowing them to co-adapt. Disabling modulation therefore does not recover the independently trained baseline in Table 1. We have clarified this distinction in the main text and both table captions.

Reviewer_tSxDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HashMod, a lightweight plug-and-play refinement module for feed-forward 3D Gaussian Splatting. It uses patch-wise binary hash matching to adaptively select multi-level backbone features and predict corrections to Gaussian parameters. Experiments across four architectures and multiple benchmarks demonstrate modest reconstruction improvements and reduced fine-tuning time. Additional evaluations on two self-collected crop datasets demonstrate its practical applicability.

Critical strengths
  1. Practical and lightweight design.

HashMod can be integrated into multiple feed-forward 3DGS architectures with relatively few additional parameters. Its ability to achieve comparable reconstruction quality with fewer fine-tuning iterations demonstrates practical value.

  1. Broad experimental evaluation.

Experiments across four architectures, multiple public benchmarks, and two self-collected agricultural datasets provide useful evidence of the method's applicability to different reconstruction scenarios.

Critical weaknesses
  1. Limited performance improvements.

Although HashMod improves reconstruction quality in most settings, the gains are generally modest and occasionally negative. The improvements over conventional FPN fusion are particularly small, making the practical advantage of hash-guided selection unclear.

  1. Insufficient validation of the core mechanism.

The experiments do not adequately establish why binary hash matching and adaptive feature selection are necessary compared with simpler alternatives, such as learned gating or attention-based fusion. More targeted comparisons are needed to demonstrate their specific advantages in cross-domain reconstruction, training efficiency, or challenging scenes.

Questions for authors

Could the authors compare HashMod with parameter-matched adaptive gating or attention-based feature selection to isolate the contribution of hash-guided matching?

Can the authors provide stronger evidence that hash-guided selection offers specific advantages in cross-domain generalization, fine-structure reconstruction, or convergence efficiency?

Could the authors provide training curves and equal-iteration comparisons to better substantiate the reported efficiency improvements?

1 comment

Thank you for your thoughtful review and recognition of HashMod’s practical value. We have added routing comparisons, training/test curves, and selection statistics to address your concerns. We hope this evidence and the clarifications below support reconsideration of your assessment.

Critical Weaknesses:

W1 Limited performance improvements

Thank you for this observation. The final gains are modest in some settings, particularly relative to FPN fusion. We have expanded the evaluation to examine reconstruction quality throughout training, alongside the existing cross-domain results. The additional evidence is discussed in our responses to Questions 2 and 3 below. We will clarify that the magnitude of improvement depends on the architecture, dataset, and metric.

W2 Validation of the core mechanism Thank you for this suggestion. We have added comparisons with continuous cosine matching, parameter-matched adaptive gating, and a variant retaining the same refinement heads without selection. We also report the distribution of selected level counts during training. These analyses examine the contribution of selection beyond the additional heads and compare alternative routing mechanisms. The settings and findings are detailed in our responses below, with convergence curves in Appendix A.3 (Training Convergence and Equal-Iteration Comparisons), routing baselines in Appendix A.4 (Comparison with Alternative Routing Mechanisms), the same-head ablation in Appendix A.5 (Same Refinement Heads without Selection), and selection statistics in Appendix A.6 (Distribution of Selected Level Counts).

Questions For Authors:

Q1 Comparison with adaptive gating

Thank you for this suggestion. We added a parameter-matched adaptive gating baseline with the same four refinement heads and routing-network parameter count as HashMod. It replaces binary matching and discrete selection with continuous matching and independent sigmoid gates, preserving multiplicative refinement.

HashMod achieves lower training reconstruction loss at every recorded epoch, with comparable final reconstruction quality. At epoch 10, its PSNR/SSIM/LPIPS are 21.535/0.733/0.300, versus 21.526/0.733/0.301 for gating. Implementation and final results appear in Appendix A.4, Table A2; training and test curves are in Appendix A.3, Fig. A2.

Q2 Evidence for selection benefits

Thank you for this suggestion.We have added supplementary experiments to address this concern.The selector does not collapse to using all four levels. During steps 95K–100K, 16.840% of patches select two levels and 83.160% select three, averaging 2.832 levels. The same-head baseline without selection achieves similar final quality but higher training reconstruction loss at every recorded epoch, supporting selection’s role in training optimization. Q3 provides matched-checkpoint comparisons.

Cross-dataset results further support the complete module’s transferability. In Table 2, models trained only on DL3DV improve PSNR, SSIM, and LPIPS with HashMod across AnySplat, YoNoSplat, and VGGT-Ω on both Tanks and Temples and Mip-NeRF 360, without target-domain fine-tuning. YoNoSplat gains 0.780 and 0.670 dB PSNR, respectively. Selection evidence appears in Appendix A.5, Table A3 and Appendix A.6, Fig. A3 and Table A4; convergence curves are in Appendix A.3, Fig. A2.

Q3 Training curves and equal iteration comparisons

Thank you for this suggestion. We provide training-loss and test-metric curves for Vanilla DA3, HashMod, adaptive gating, and cosine matching, evaluated at matched training iterations. These supplementary results come from separate training runs from those reported in Table 1. Each epoch in the supplementary experiments contains approximately 10K iterations, giving approximately 100K iterations over 10 epochs. In these supplementary runs, HashMod reaches competitive reconstruction quality earlier. At epoch 3, it achieves 21.366 dB PSNR, 0.730 SSIM, and 0.298 LPIPS, exceeding the Vanilla DA3 epoch-10 results of 21.253 dB, 0.725, and 0.303. At epoch 8, HashMod achieves 21.516 dB, 0.732, and 0.301, achieving higher PSNR with comparable SSIM and LPIPS to cosine matching's epoch-10 results of 21.485 dB, 0.732, and 0.301. Relative to adaptive gating, HashMod maintains lower training reconstruction loss and comparable final reconstruction quality. These curves show that HashMod reaches competitive reconstruction quality with fewer training iterations and maintains lower training reconstruction loss than the compared alternatives at matched iterations. This provides further evidence of improved convergence efficiency. The complete training-loss and test-metric curves are provided in Appendix A.3 (Training Convergence and Equal-Iteration Comparisons), Fig. A2.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.