acceptodds
Under review as a conference paper at ICLR 2027

READ: Geometry-Guided Source Evidence Retrieval for Feed-Forward 3D Gaussian Splatting

Abstract

Feed-forward 3D Gaussian splatting enables efficient novel view synthesis, but fine appearance details can be lost when source observations are encoded into Gaussian attributes. We present Reprojected Evidence Aggregation via Depth (READ), a lightweight rendering-time refinement module that uses a frozen Gaussian scene for both base rendering and geometry-guided retrieval of source appearance. READ combines posed source images with rendered RGB, depth, and accumulated alpha, without accessing the reconstruction model's internal features. Geometric reliability and learned source weights guide a bounded residual correction of the base rendering. With only 21K trainable parameters, READ improves novel-view fidelity and visual detail on DL3DV and RealEstate10K with low computational overhead. A single set of weights trained with one reconstruction model transfers across Gaussian reconstruction models, source-view counts, and datasets without retraining.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
novel view synthesis, 3D Gaussian Splatting, feed-forward reconstruction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_EMuwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this work, the authors propose READ, a lightweight rendering-time refinement module for pre-trained frozen feed-forward 3DGS prediction networks. Specifically, READ queries the input source views by target-to-source reprojection and uses the queried evidence to predict a bounded residual, which is introduced to refine the rendered target view. According to the experiments on DL3DV and Re10k, READ can enhance the rendering quality with low additional computation cost and zero-shot transferring ability.

Critical strengths
  1. The idea to refine novel view synthesis with an extra network is reasonable;

  2. The presentation is clear and easy to follow, where the experiments also show clear improvements.

Critical weaknesses
  1. Some ablation experiments are missing. Since the main contribution of this work is the construction of the queried evidence-to-residual network, the necessity of items in the queried evidence should be carefully analyzed, while existing experiments have not clearly illustrate the effectiveness of, e.g., FiF_i and ϕi\phi_i;
Questions for authors
  1. As mentioned in the weakness section, some extra ablation experiments may be needed;

  2. Some discussions about [1] could be included, which is also an enhancement approach for a initial 3DGS object;

[1] Splatformer: Point Transformer for Robust 3D Gaussian Splatting

  1. The setting of this work is constrained to 4/8 source views-conditioned 3DGS feed-forward prediction frameworks. Is it possible to support continuous partial view inputs like [2] instead of whole source view inputs?

[2] FreeSplat: Generalizable 3D Gaussian Splatting Towards Free-View Synthesis of Indoor Scenes.

  1. This method may be incompatible with a 3DGS scene without source images. Is it possible to introduce image generation/editing models to enhance this method?
1 comment
AuthorsOpenReview ↗

Response to EMuw: Ablations, Related Work, and Partial-Input Generalization

Thank you for the helpful questions. The revision includes the requested evidence ablations, SplatFormer discussion, and FreeSplat partial-input experiments.

W1/Q1: shallow source features FiF_i and relative-view information ϕi\phi_i.

Table 8 compares the requested ablations and controls for evidence processing and source selection on DL3DV. Δ\DeltaPSNR is measured relative to TokenGS Base:

VariantPSNR ↑SSIM ↑LPIPS ↓ΔPSNR ↑
TokenGS Base22.7920.7320.331—
GeometryBlend22.8120.7270.291+0.020
Uniform weighting23.6250.7640.262+0.833
Fixed reliability weighting23.6180.7640.263+0.826
Without FiF_i23.7810.7670.256+0.989
Without ϕi\phi_i23.6630.7640.260+0.871
Full READ23.9300.7710.249+1.138

Removing either FiF_i or ϕi\phi_i reduces all three quality metrics. FiF_i supplies local source-image structure complementary to RGB, while ϕi\phi_i encodes relative viewing directions to inform source selection. Their gains support complementary roles in representing and selecting source evidence.

The other controls clarify the broader design. GeometryBlend uses fixed RGB fusion and yields little PSNR gain. Uniform/Fixed weighting adds learned evidence processing and substantially improves quality; Full READ further improves all three metrics through target-conditioned source selection. Thus, geometric retrieval provides the observations, while learned processing and selection turn them into effective refinement.

Q2: Relation to SplatFormer.

Section 2.1 now discusses SplatFormer [1]: its point transformer updates Gaussian attributes to improve the scene representation used for rendering. READ keeps the Gaussian scene fixed and retrieves source evidence to correct each target image. Its RGB/depth/alpha interface allows the same refinement network to operate across different reconstruction architectures.

Q3: continuous partial inputs and FreeSplat.

We evaluate FreeSplat-fv [2] on Replica with 2/4/6/8/10 arriving source views on fixed targets, using READ weights trained with TokenGS on DL3DV. Mean PSNR gains over FreeSplat-fv are shown below; Table 21 provides the absolute scores:

Arrived source views246810
Δ\DeltaPSNR ↑ (dB)+0.061+0.566+1.049+1.629+2.001

Mean gains are positive at every tested prefix, with a small gain at two inputs. READ also improves FreeSplat when newly arriving views update an existing Gaussian scene, providing a 2.253 dB PSNR gain at ten input views (Table 22).

These FreeSplat experiments directly address continuous partial inputs: READ can enhance novel views as source images arrive, after either reconstruction from the available views or incremental scene updates, using the same TokenGS-trained weights throughout. It therefore does not require the full source set in advance. Transfer from TokenGS on DL3DV to FreeSplat on Replica also demonstrates generalization to a new reconstruction model and dataset without retraining READ.

Q4: combining READ with image generation or editing.

Thank you for suggesting this promising direction. READ is designed as an add-on that recovers detail from available source observations with minimal additional latency. Its interface takes source images, their cameras, and rendered scene geometry, providing a natural connection to a source-view generation model. For scenes whose original source images are unavailable, such a model could generate replacement source views conditioned on the existing scene and specified cameras. The generator would provide appearance details, while READ would retain its role of using scene geometry to retrieve and fuse them into target renderings. We highlight this integration and the study of cross-view consistency in generated evidence as future work in Section 4.5.

[1] Chen et al., SplatFormer, ICLR, 2025.

[2] Wang et al., FreeSplat, NeurIPS, 2024.

Reviewer_yLjWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces READ, an add-on module for feed-forward 3D Gaussian prediction which helps address the loss of fine details through a lightweight 21K parameter module. Given a train 3D Gaussian scene, this technique uses its rendering from a novel view along with depth and CNN features to predict a delta to be applied to the rendered image. The READ module is trained once and then can be applied across datasets and models. Results on DL3DV and RE10K show that READ helps improve the novel view quality with minimal additional overhead at inference time.

Critical strengths
  1. The READ module is light-weight and efficient while staying performant. In addition, it shows strong generalization, as a model trained on DL3DV generalizes well to RE10K and the technique shows improvement for both DepthSplat and TokenGS.
  2. The architecture is thoughtfully designed to limit any negative effects to gating the updates with depth consistency and opacity
Critical weaknesses
  1. The proposed technique is strongly dependent on the spacing and quality of input source views. As shown in Fig. 5, the majority of READs improvement comes from situations in which the novel view is quite close a source view, since in these cases the depth reprojection contains valuable information. The PSNR gain drops sharply as we move away from a source view, and even goes close 0 when we are not near any source view.
  2. The related works section introduces other refinement/source view aggregation techniques, but these are not compared against in the paper, limiting the confidence in the quality of READ as compared to prior work.
Questions for authors
  1. Given that READ requires source images, camera poses, feature maps, and more at inference time, what are the memory requirements and how much additional overhead is there on top of the base model?
  2. In sparse-view settings, there may be target regions that are largely or entirely not covered in source images. Does READ degrade gracefully in these conditions?
5 comments
AuthorsOpenReview ↗

Response to yLjW: Evidence Fusion, Source Support, and Efficiency

Thank you for the constructive review. We address each concern below.

W1: Dependence on source distance.

On DL3DV, we report PSNR changes and LPIPS reductions relative to TokenGS for READ and MaRINeR [1], grouping targets from nearest to farthest within each scene (Appendix D.1):

Source-distance group1 (nearest)234 (farthest)
READ ΔPSNR (dB) ↑+2.635+0.999+0.587+0.247
MaRINeR ΔPSNR (dB) ↑−1.638−2.142−2.461−2.455
READ LPIPS reduction ↑0.13320.08600.06310.0397
MaRINeR LPIPS reduction ↑0.13700.10500.08120.0584

READ retains a positive mean gain in the farthest group (95% CI: [+0.183, +0.322] dB). MaRINeR, which uses source images for enhancement, also shows decreasing LPIPS gains with distance, reflecting the common dependence of source-guided enhancement on relevant observations. We discuss its LPIPS–PSNR tradeoff in W2. READ benefits most from nearby sources but retains positive mean gains farther away.

W2: Contribution and comparison with existing enhancement methods.

We compare READ with pretrained MaRINeR and the Difix [2] inference-time enhancers on TokenGS renderings from DL3DV (Table 5). MaRINeR uses feature matching to transfer reference-image detail, while Difix and Difix_ref use diffusion-based enhancement without and with a reference image, respectively. READ instead uses existing Gaussian geometry for source retrieval, allowing its 21K-parameter network to focus on target-conditioned evidence fusion:

MethodPSNR ↑SSIM ↑LPIPS ↓Time (ms) ↓Extra time (ms) ↓
TokenGS22.7920.7320.330650.20—
TokenGS + Difix21.2030.6620.2987131.3181.13
TokenGS + Difix_ref21.5320.6800.2515152.43102.32
TokenGS + MaRINeR20.6230.6900.2346135.3585.99
TokenGS + READ23.9300.7710.249355.565.41

Times are measured end-to-end on an A100; extra time denotes the median paired increase (Appendix A.3).

READ achieves the highest PSNR/SSIM and lowest latency among these enhancers. MaRINeR's lower LPIPS is consistent with its perceptual and adversarial training objectives. These comparisons address both detail recovery and efficiency: MaRINeR and Difix add substantial runtime to the base renderer, while READ improves all three quality metrics with a small additional cost. The results support our central insight: using existing geometry for retrieval lets a compact network recover detail while preserving the speed advantage of feed-forward 3DGS. The ablations in Table 8 further clarify the role of learning after geometric retrieval: learned evidence processing provides the main PSNR gain over GeometryBlend, and adaptive source selection adds about 0.3 dB over uniform or fixed weighting.

Q1: Memory and additional overhead.

READ adds 0.115 GiB of peak allocated memory and 5.41 ms per target for four-view TokenGS on DL3DV (A100, 448×256), including source depth/alpha rendering. More or higher-resolution source images require more memory for their RGB and feature maps, while the same network processes every source with a fixed parameter count. Reusing depth/alpha maps across targets for a fixed scene and source set amortizes their rendering cost. Details are in Appendix A.3.

Q2: Limited or absent source coverage.

For four-view TokenGS on DL3DV, we group target pixels by source support estimated from predicted geometry; absent coverage means outside all source frusta. All method columns report RGB MSE ↓ in [0,1]:

Source supportPixels (%)TokenGS+READ+Difix+MaRINeR
Multiple sources (≥2)50.650.004800.003880.007140.00977
Limited support (1 source)31.300.007850.007280.010600.01366
Covered, but depth-inconsistent17.450.013520.013400.017050.02282
Absent coverage0.600.017060.016960.020620.02524

READ reduces average error with single-source support. Under depth-inconsistent or absent coverage, gains are smaller, but MSE remains below TokenGS; Difix and MaRINeR produce higher MSE. Measured mean absolute RGB changes under absent coverage are about 16% of those in multi-source regions (Appendix D.2), supporting stable average reconstruction quality as source support weakens.

This RGB-error comparison complements the perceptual gains of these enhancers in W2; lower MSE and lower LPIPS measure different aspects of quality. Since most evaluated pixels have source support, READ recovers observed detail across much of the image while making smaller changes elsewhere. Together with its low additional latency (W2), this supports strong overall enhancement while preserving feed-forward efficiency.

[1] Bösiger et al., MaRINeR, ECCV, 2024.

[2] Wu et al., Difix3D+, CVPR, 2025.

Dear Reviewer yLjW,

Thank you once again for your thoughtful feedback and valuable suggestions. With the ICLR discussion period approaching its conclusion, we wanted to respectfully follow up and ask whether our response and additional analyses have adequately addressed the concerns raised in your review. Specifically, we added:

  1. A source-distance analysis showing that READ’s gain decreases with distance but remains positive on average in the farthest group.
  2. Matched quality–latency comparisons with MaRINeR and Difix3D+ using the same frozen TokenGS scenes and target views.
  3. Detailed measurements of READ’s additional memory and inference overhead, including source-buffer preparation and the costs that can be reused across target views.
  4. A source-coverage analysis covering multi-source, single-source, depth-inconsistent, and source-frustum-absent regions.

These results are reported in Table 5 and Appendices A.3 and D. We also revised the manuscript to state READ’s dependence on available source evidence explicitly and to clarify its quality–cost position relative to reference-image and diffusion-based enhancement methods.

If any points remain unclear or if you have any further questions or suggestions, we would be very grateful for the opportunity to discuss them with you before the discussion deadline. Thank you very much for your consideration.

Best regards,
The Authors

Dear Reviewer, The authors have posted their rebuttal responses. I would kindly encourage you to review the authors' responses and continue the discussion by providing folow-up comments where appropriate. In particular please indicate whether the authors responses have adequately addressed your concerns and update your assessment if necessary Thank you for your time and continued efforts! Your AC

Reviewer_NEkXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

READ introduces a lightweight module for refining novel view renderings from feed-forward 3D Gaussian reconstruction models. It uses rendered depth to align source colors and shallow features with the target view, combines them according to geometric reliability and learned weights, and predicts a bounded image correction while keeping the Gaussian scene frozen. The module has approximately 21K trainable parameters, and the same weights trained with TokenGS are reused across the evaluated reconstruction models and settings.

The design is reasonable, and the reported quantitative and visual results show useful improvements with low additional latency. However, the methodological contribution is incremental, primarily integrating established reprojection, feature fusion, and residual correction techniques into a lightweight module. Its advantage over existing rendering enhancement methods is not established through direct comparisons. Evaluation also has limited coverage of recent reconstruction models and is restricted to sparse source views, leaving its broader practical value insufficiently demonstrated. My current recommendation is weak rejection.

Critical strengths
  1. Reasonable and lightweight method. Recovering details from geometrically aligned source observations directly addresses information lost in Gaussian appearance prediction. Reliability weighting and bounded residual correction form a coherent design.

  2. Transfer across the evaluated models. The same 21K-parameter network improves TokenGS, DepthSplat, and ReSplat without retraining. The experiments also demonstrate transfer across input view counts and resolutions, with additional cross-dataset results using TokenGS.

  3. Good experimental results with low overhead. The quantitative results show clear improvements over the corresponding base models, and the displayed images recover sharper textures and small details. In the four-view TokenGS setting, PSNR increases from 22.792 to 23.930 dB while the reported single-target runtime increases from 50 to 56 ms.

Critical weaknesses
  1. Limited coverage of recent reconstruction models. The experiments cover TokenGS, DepthSplat, and ReSplat, but omit recent models such as YoNoSplat, tttLRM, and WorldMirror. The benefit of READ across a broader range of current reconstruction systems remains insufficiently established.

  2. Evaluation is restricted to sparse inputs. All experiments use two to eight source views. It remains unclear whether READ provides worthwhile quality gains and low overhead with denser source coverage.

  3. Missing comparison with rendering repair methods. Although Difix3D+ and other enhancement methods are discussed, the experiments focus on base renderers and token tuning. Direct comparisons are needed to establish READ's relative image quality and computational cost.

Questions for authors
  1. Could the authors evaluate READ on a representative recent reconstruction model, such as YoNoSplat, tttLRM, or WorldMirror, and report its additional quality gains and runtime?

  2. How does READ perform with denser source coverage, using a reconstruction model that supports larger input sets? Does its quality improvement remain worthwhile relative to the additional rendering and retrieval cost?

  3. Could the authors compare READ with a representative rendering repair method, such as the inference-time enhancement stage of Difix3D+, using the same frozen scene and target views, and report image quality and latency?

8 comments
AuthorsOpenReview ↗

Response to NEkX: Generalization, Dense Inputs, and Efficient Enhancement

Thank you for your review. We address your concerns below.

W1/Q1: Transfer to recent reconstruction models.

The unchanged TokenGS-trained READ weights improve all three quality metrics for WorldMirror and YoNoSplat on DL3DV (Table 6):

Model / viewsResolution (W×H)MethodPSNR ↑SSIM ↑LPIPS ↓Time (ms) ↓Extra time (ms) ↓
WorldMirror / 8518×378Base21.5720.6880.1931029.49—
+READ24.3140.7650.1421054.9724.50
YoNoSplat / 6224×224Base24.7580.8180.139106.27—
+READ26.4590.8460.124116.109.47

Times are measured end-to-end on an A100; extra time denotes the median paired increase (Appendix A.3).

READ retrieves source evidence using rendered RGB, depth, and alpha maps available across these reconstruction models. This shared interface allows the same refinement network to improve different backbones without model-specific adaptation, demonstrating the generalization value of separating evidence fusion from reconstruction.

The same READ weights also improve FreeSplat-fv [1] on Replica with ten input views across all three metrics (Table 7), demonstrating transfer to both a different reconstruction architecture and a new dataset.

W2/Q2: Quality and cost with denser source inputs.

Following the YoNoSplat paper's DL3DV evaluation settings, we cover all three reported source-view configurations, from sparse to dense inputs: 6, 12, and 24 views. The same READ weights improve all three quality metrics in every configuration (Tables 6 and 12):

Model / viewsMethodPSNR ↑SSIM ↑LPIPS ↓Time (ms) ↓Extra time (ms) ↓
YoNoSplat / 6Base24.7580.8180.139106.27—
+READ26.4590.8460.124116.109.47
YoNoSplat / 12Base23.3120.7740.177156.49—
+READ26.1660.8270.145170.7614.67
YoNoSplat / 24Base22.2060.7420.197280.40—
+READ24.7280.8050.163315.3634.49

Times are for one target view and include source depth/alpha rendering.

READ improves quality over the corresponding base at every tested source count, including 24-view YoNoSplat. These gains show that increasing reconstruction inputs does not eliminate the benefit of source-guided refinement: READ retrieves source-image details that are not fully preserved in the base renderings. Additional cost grows with source count because more source depth and alpha maps must be rendered, but remains a modest fraction of YoNoSplat's inference time; these maps can also be reused across targets for a fixed scene and source set. Together, these results support our design: Gaussian geometry supplies correspondences, allowing a compact network to focus on selecting and fusing source evidence, with quality gains and low additional cost even under dense inputs.

W3/Q3: Rendering-enhancement quality and latency.

We compare READ with MaRINeR [2] and the Difix [3] inference-time enhancers on TokenGS renderings from DL3DV (Table 5). MaRINeR uses feature matching to transfer reference-image detail, whereas Difix and Difix_ref use diffusion-based enhancement without and with a reference image, respectively:

MethodPSNR ↑SSIM ↑LPIPS ↓Time (ms) ↓Extra time (ms) ↓
TokenGS22.7920.7320.330650.20—
TokenGS + Difix21.2030.6620.2987131.3181.13
TokenGS + Difix_ref21.5320.6800.2515152.43102.32
TokenGS + MaRINeR20.6230.6900.2346135.3585.99
TokenGS + READ23.9300.7710.249355.565.41

READ achieves the highest PSNR/SSIM and lowest latency among the evaluated enhancers. MaRINeR's lower LPIPS is consistent with its perceptual and adversarial training objectives. Both MaRINeR and Difix add substantial runtime to the base renderer, while READ improves all three quality metrics with a small additional cost.

Our contribution is a geometry-driven route to enhancement that addresses detail recovery and feed-forward efficiency together. Where MaRINeR estimates feature correspondences and Difix performs diffusion-based refinement, READ reuses the reconstructed 3DGS geometry to retrieve source detail and assigns evidence selection and fusion to a shared 21K-parameter network. Scene geometry supplies correspondences, while learning focuses on appearance correction. The resulting substantial quality gains at low additional cost, together with fixed-weight transfer across evaluated models and input counts, demonstrate the value of this design. We revised the Introduction to clarify this insight and its contributions to efficiency and transferability.

[1] Wang et al., FreeSplat, NeurIPS, 2024.

[2] Bösiger et al., MaRINeR, ECCV, 2024.

[3] Wu et al., Difix3D+, CVPR, 2025.

Dear Reviewer NEkX,

We sincerely appreciate your careful review and constructive suggestions. As the ICLR discussion period is drawing to a close, we wanted to kindly follow up to ask whether our response and additional experiments have adequately addressed the concerns raised in your review. Specifically, we added:

  1. Evaluations on recent reconstruction models, including WorldMirror and YoNoSplat, together with cross-model and cross-dataset transfer to FreeSplat on Replica.
  2. Dense-input experiments with 6, 12, and 24 source views, showing consistent improvements across all three quality metrics.
  3. Matched quality–latency comparisons with MaRINeR and Difix3D+ using the same frozen TokenGS scenes and target views.

These results are reported in Tables 5–7 and 12 and Appendix A.3. We also revised the Introduction to clarify READ’s central contribution: it reuses reconstructed Gaussian geometry to retrieve and aggregate source evidence with a shared 21K-parameter network, while introducing substantially less additional latency than the other evaluated enhancement methods.

If any points remain unclear or if you have any further questions or suggestions, we would be very grateful for the opportunity to discuss them with you before the discussion deadline. Thank you very much for your consideration.

Best regards,
The Authors

Thank you to the authors for providing the additional experiments. However, the original YoNoSplat paper reports post-optimization results that appear to be generally better than the additional YoNoSplat + READ results presented in the rebuttal. Given this comparison, could the authors clarify what they consider to be the main contribution and advantage of READ?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Our revisions summary

We thank the reviewers for their constructive feedback. We first highlight the motivation and central insight of READ. Existing methods for enhancing 3DGS rendering quality can add substantial latency, diminishing the efficiency advantage of feed-forward 3DGS. READ instead uses reconstructed geometry for source retrieval, allowing a small shared network to focus on evidence selection and fusion. This recovers detail at low additional cost and supports fixed-weight transfer across reconstruction models and input settings. Following the reviewers’ suggestions, we added experiments and expanded the discussion, as summarized below.

  1. Enhancement quality and computational cost (NEkX, yLjW). Comparisons with MaRINeR and the Difix inference-time enhancers show that READ achieves the highest PSNR/SSIM and lowest latency among the evaluated enhancers (with only 5–7% of their additional runtime), while substantially improving LPIPS over the base renderer (Table 5). We also added memory measurements and paired runtime comparisons across recent reconstruction models (Table 6; Appendix A.3).

  2. Generalization across models and input settings (NEkX, EMuw). The same TokenGS-trained weights improve WorldMirror, YoNoSplat, and FreeSplat on Replica, demonstrating cross-model and cross-dataset transfer (Tables 6–7). READ substantially improves all three quality metrics across YoNoSplat’s 6-, 12-, and 24-view configurations with modest additional cost. FreeSplat experiments further demonstrate enhancement as source views arrive and after incremental Gaussian-state updates (Appendix D.3).

  3. Source distance and coverage (yLjW). READ retains positive mean gains in the farthest source-distance group; MaRINeR also shows decreasing perceptual gains with distance (Appendix D.1). Regional analysis supports READ’s design choice: making smaller changes where evidence is weak and improving quality in source-supported regions, which account for most evaluated pixels. This yields substantial overall gains while maintaining stable average quality under limited or absent coverage (Appendix D.2).

  4. Evidence ablations (EMuw). We added the requested source-feature and relative-view ablations. Controls using fixed RGB fusion and different weighting schemes further establish the roles of learned evidence processing and adaptive source selection (Table 8).

  5. Related work and future directions (NEkX, EMuw). We expanded comparisons with existing enhancement approaches and distinguished SplatFormer’s Gaussian-attribute refinement from READ’s target-image refinement (Section 2). Integration with source-view generation is discussed as future work for scenes whose original source images are unavailable (Section 4.5).

We thank all three reviewers for their thoughtful engagement and constructive feedback, which have helped us strengthen the evaluation and presentation of our work. Reviewers NEkX and yLjW have confirmed that our responses addressed their main concerns and have raised their ratings, and we appreciate Reviewer EMuw’s positive feedback. The follow-up experiments further demonstrate that READ can complement post-optimization, reinforcing its value as a lightweight, transferable rendering-enhancement module.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.