acceptodds
Under review as a conference paper at ICLR 2027

ImaGeo: Bidirectional Image–Geometry Co-Reasoning for 3D Generation

Abstract

Image-conditioned 3D generation requires fidelity to a single image’s visible structure and details, and consistency with complementary observations when multiple views are available. We present ImaGeo, a diffusion transformer that generates 3D geometry from one or more input images. Its denoising blocks combine per-view image and geometry self-attention with joint attention, connecting all views to a shared geometry representation. Updated image features carry geometric and cross-view context into subsequent blocks while retaining per-view spatial processing. The same design operates in coarse and fine generation, trained solely through geometry flow matching without a dedicated reconstruction branch or auxiliary reconstruction losses. It supports variable-view inputs without input camera parameters. On GSO and Toys4K, ImaGeo improves single-view image–geometry consistency and sparse-view reconstruction accuracy. Evaluations on AI-generated images also show improved image–shape similarity. In controlled single-view fine-stage comparisons, our 1.5B denoiser outperforms a parameter-matched cross-attention baseline and achieves normal-map quality comparable to a 5B baseline.

Primary area:
generative models
Keywords:
3d generation, mlulti-view 3D reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_B28qOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes ImaGeo, a general-purpose model for single- and multi-view image-to-3D generation. It introduces an attention architecture in which image and geometry tokens are jointly updated and interact with each other, allowing image tokens to evolve during generation and support geometry prediction in the next layer.

Critical strengths
  1. Simple yet effective. ImaGeo achieves state-of-the-art performance using bidirectional attention and 325K training assets, fewer than several competing methods. The provided qualitative comparisons also show that ImaGeo generates more accurate geometry with richer details than other methods.
Critical weaknesses
  1. Insufficient experimental evaluation. The paper acknowledges that similar image–shape co-reasoning is used in Hunyuan3D 2.0, so a controlled comparison with its architecture trained on the same data should be included. Moreover, replacing cross-attention with joint attention is expected to substantially increase computational cost, with the gap becoming more pronounced as the number of input images increases. The paper should quantify and report these costs alongside the final generation quality.
Questions for authors
  1. How much slower is ImaGeo than the cross-attention baseline? Does this gap widen as the number of input images increases?
  2. Could the authors provide more details on data selection and processing? The paper only states that 325K assets were selected from Objaverse-XL. How exactly were these assets selected and processed? Despite using fewer training assets than TRELLIS and LATTICE, ImaGeo shows clear performance gains, suggesting that data selection and processing may contribute substantially beyond the architectural changes.
3 comments

Response to Reviewer B28q

Dear Reviewer B28q,

Thank you for your careful review and helpful suggestions. Below, we address your concerns about related architectures, computational cost, and data preparation through the revised experiments and a more detailed description of our data processing pipeline.

Q1. Architecture comparisons

Since related image–shape interaction already exists, how is the architectural contribution established without a Hunyuan-style model trained on the same data?

Response. We add an MMDiT-style + view-ID baseline that removes per-view image self-attention while retaining shared-QKV joint attention (Section 4.4, Table 5). View-ID modulation distinguishes tokens from different views after per-view processing is removed.

With four views and the same Stage 2 model, the Table 3 ImaGeo checkpoint outperforms this Hunyuan/MM-DiT-inspired variant on all five metrics—CD-L1, CD-L2, and normal-map PSNR, SSIM, and LPIPS—on both GSO and Toys4K. For example, PSNR improves by 0.479/0.351 dB, respectively.

Q2. Efficiency and view-count scaling

How much slower is ImaGeo than cross-attention, and how does the cost gap change with the number of input views?

Response. Table 4 shows gains over CA 1.5B on all four quality metrics and comparable quality to CA 5B. Tables 6–7 report Stage 2 DiT costs on one L20 (FP16, 32,768 geometry tokens). With one view, ImaGeo uses only 6.0% more FLOPs and 2.9% more CFG-step time than CA 1.5B, but 46.8% fewer FLOPs and 46.7% less time than CA 5B. Peak allocated memory is 7.756 GiB, versus 8.915/19.563 GiB for CA 1.5B/5B.

From one to six views, ImaGeo’s CFG-step latency increases from 7.855 to 15.969 s. The multi-view cost gap relative to CA remains unmeasured.

Q3. Data selection and preprocessing

How were the 325K assets selected and processed, and could data curation explain the gains beyond the architecture?

Response. We have added the data selection and preprocessing details to Appendix A.4. After initial filtering (~500K assets), we apply Sparc3D-based flood-fill reconstruction [1] and reject meshes with more than 15% internal vertices, retaining ~325K training assets. All architectures in Table 4 use the same training data and preprocessing, controlling for data curation in this comparison.

[1] Zhihao Li, Yufei Wang, Heliang Zheng, Yihao Luo, and Bihan Wen. Sparc3D: Sparse Representation and Construction for High-Resolution 3D Shapes Modeling. NeurIPS, 2025.


Thank you again for your time and consideration. We hope these clarifications and additional results address your concerns. We welcome any further questions and look forward to your feedback.

Reviewer_B28qOpenReview ↗

Thanks for the clarification

Thank you for the detailed response and additional results. My concerns have been addressed, and I have raised my score accordingly.

Thank you for carefully considering our responses and additional results. We are glad that your concerns have been addressed, and we appreciate your constructive feedback, which helped us improve the paper.

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

This paper proposes ImaGeo, a diffusion transformer for image-conditioned 3D geometry generation, built on LATTICE’s two-stage (VecSet coarse, VoxSet fine) representation. Instead of geometry tokens cross-attending to fixed image features, each block applies per-view image and geometry self-attention, followed by joint attention over the timestep, geometry and all view tokens. The updated image tokens are passed to the next block. The model is trained only with a geometry flow-matching loss and handles 1–6 views without camera parameters or auxiliary reconstruction losses. It is evaluated on single-view generation and 4-view reconstruction, and in a 40-object ablation where the 1.5B model is compared against 1.5B and 5B cross-attention baselines. The authors report the best normal-map and Chamfer metrics across baselines, and the 1.5B model is roughly on par with the 5B baseline.

Critical strengths
  1. Well-motivated design. Updating image tokens through joint attention, rather than keeping them fixed as in cross-attention, is a reasonable and practical change. It supports 1-6 views with no camera input and no auxiliary losses.
  2. Fairly controlled comparison. Table 4 fixes the VAE, data, Stage 1 outputs and inference settings, and compares against both 1.5B and 5B cross-attention baselines. The paper discloses training budgets, the GT-assisted alignment caveat and failure cases, which makes the claims checkable.
Critical weaknesses
  1. Table 4 compares the full ImaGeo block with a cross-attention block that differs in many other ways: depth, extra MLPs, prefix tokens and shared-QKV joint attention. There is no ablation that freezes image tokens while keeping joint attention, and no MM-DiT or Hunyuan3D-style baseline. The gains are small, on 40 objects with no variance estimates.

  2. The cross-attention baseline also never appears in the main benchmark tables (Tables 1 and 3), so the headline gains cannot be attributed to IGCA rather than the LATTICE backbone, the data or the protocol.

Questions for authors

I would like the authors to address the three weaknesses above.

3 comments

Response to Reviewer BRCz

Dear Reviewer BRCz,

Thank you for your thoughtful review and helpful suggestions on isolating the role of image–geometry interaction and strengthening the comparisons. We have added targeted ablations, uncertainty estimates, and a cross-attention baseline, and address your concerns point by point below.

Q1. Attributing gains to image–geometry interaction

How can the benefit be attributed to image–geometry interaction when Table 4 changes several block components, and what about freezing image tokens while retaining joint attention?

Response. Table 4 compares parameter-matched 1.5B models, with only 6.0% more DiT FLOPs for ImaGeo (Table 6). The Stage 1 feedback-disabled variant retains depth, shared QKV, image updates, and image-to-geometry attention (Table 5).

On 200 GSO/Toys4K objects with four views and the same Stage 2 model, the feedback-disabled variant scores below Full on all five metrics on both datasets at the evaluated checkpoints, with 58.2% higher pooled CD-L1. After 10,000 fine-tuning updates, when evaluation metrics had plateaued, it still trails Full on all five metrics, with CD-L1 4.9% higher. Frozen-image joint attention remains a separate, untested control.

Q2. Architecture comparisons

How does ImaGeo compare with MM-DiT- or Hunyuan-style architectures?

Response. We add an MMDiT-style + view-ID baseline that removes per-view image self-attention while retaining shared-QKV joint attention (Section 4.4, Table 5). View-ID modulation distinguishes tokens from different views after per-view processing is removed.

With four views and the same Stage 2 model, the Table 3 ImaGeo checkpoint outperforms this Hunyuan/MM-DiT-inspired variant on all five metrics—CD-L1, CD-L2, and normal-map PSNR, SSIM, and LPIPS—on both GSO and Toys4K. For example, PSNR improves by 0.479/0.351 dB, respectively.

Q3. Statistical reliability

Are the gains in Table 4 reliable given the small evaluation set and missing variance estimates?

Response. We reran all three architectures in Table 4 on 20 Objaverse objects and 20 artist-sculpted models, with five inference seeds each (600 results). We report mean ± SD across five test-set means and paired 95% CIs from 10,000 object-bootstrap resamples after averaging seeds per object.

ImaGeo improves all four metrics—PSNR, SSIM, LPIPS, and IoU—over CA 1.5B, with all four paired CIs excluding zero. For example, PSNR improves by 1.020 dB (95% CI [0.945, 1.100]) and LPIPS decreases by 18.85%. Against CA 5B, observed quality is comparable, and all four CIs include zero.

Q4. Controlled cross-attention baselines

Without the controlled cross-attention baseline in Tables 1 and 3, can the headline benchmark gains be attributed specifically to IGCA?

Response. We have added a 1.5B cross-attention baseline to Table 1. ImaGeo improves all three normal-map metrics—PSNR, SSIM, and LPIPS—on both GSO and Toys4K. For example, PSNR increases by 0.712/1.108 dB, respectively.

For multi-view reconstruction, we expect a naive CA extension with fixed, independently encoded image features to be less effective because it lacks explicit cross-view refinement. We therefore prioritized the Hunyuan/MM-DiT-inspired comparison in Table 5, where ImaGeo improves all five metrics on both datasets, instead of training a multi-view CA baseline.


Thank you again for your time and consideration. We hope these clarifications and additional results address your concerns. We welcome any further questions and look forward to your feedback.

Thank you for the authors' responses. The additional experiments have addressed most of my concerns. I have no further questions. I hope these experiments can be included in the final version of the paper. I will keep my current score.

Thank you for your constructive feedback and for considering our responses. The additional experiments are already included in the revised PDF uploaded to OpenReview, and we will retain them in the final version. We appreciate your time and helpful suggestions.

Reviewer_z7o6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes ImaGeo, a diffusion transformer that generates 3D geometry from one or more uncalibrated images. Each denoising block applies per-view image self-attention and geometry self-attention, then a joint attention over all image and geometry tokens. The updated image tokens are carried into the next block, rather than kept as fixed cross-attention context. The design is used in both stages of LATTICE’s coarse-to-fine VecSet/VoxSet pipeline and trained only with geometry flow matching, without reconstruction branches or auxiliary losses. Trained on 325K Objaverse-XL assets, ImaGeo reports the best single-view and four-view results on GSO and Toys4K. Its 1.5B Stage 2 denoiser outperforms a parameter-matched cross-attention baseline and approaches a 5B one.

Critical strengths
  1. The paper is clearly written, and the method is simple. One block design handles single-view generation and sparse-view reconstruction with a variable number of views, needs no input cameras, and uses no supervision beyond geometry flow matching.

  2. The paper includes a controlled architecture comparison against parameter-matched and 5B cross-attention baselines that share the VAE, data, curriculum and inference settings (Table 4). Prior joint image–shape designs such as Hunyuan3D 2.0 did not report this kind of comparison.

Critical weaknesses
  1. The architectural novelty is limited, and the central “bidirectional co-reasoning” claim is not isolated. The block combines MM-DiT-style joint attention, already used for image–shape tokens in Hunyuan3D 2.0, with the per-view/global attention pattern of VGGT and π³. IGCA also shares its QKV projections across modalities and uses no positional encoding (A.1), so it is plain attention over concatenated tokens. No experiment shows that geometry-to-image feedback is what helps. There is no variant in which image tokens are refined per block without attending to geometry, no comparison with Hunyuan3D-style dual/single-stream blocks, and no analysis of what the image tokens learn from geometry.

  2. The controlled comparison is thin and not compute-matched. Table 4 uses 40 objects, partly from a private dataset, in a single run without variance. It covers only Stage 2, with a single view and ground-truth Stage 1 inputs. The gains over the 1.5B baseline are small (+0.31 dB PSNR, +0.002 IoU). Matching parameters does not match compute. ImaGeo carries 5,329 image tokens per view through self-attention, joint attention and MLPs in all 21 blocks, while the baseline uses its image features only as cross-attention context. No FLOPs, latency or memory are reported, so “comparable to a 5B baseline” may reflect extra image processing rather than a more efficient design. The multi-view case, where joint attention is claimed to integrate views, has no architecture comparison at all.

Questions for authors
  1. Can you ablate a variant in which image tokens are updated each block by self-attention only, without attending to geometry, to isolate the benefit of bidirectional interaction? How does your block compare with MM-DiT-style modality-specific QKV and with Hunyuan3D 2.0’s dual/single-stream design?

  2. Can you report FLOPs, latency and memory for the 1.5B ImaGeo, 1.5B and 5B baselines at Stage 2 sequence lengths? Can you extend Table 4 to Stage 1, to multi-view inputs and to a public evaluation set, with multiple seeds?

10 comments

Response to Reviewer z7o6

Dear Reviewer z7o6,

Thank you for your constructive feedback. We have clarified our contribution and added experiments to address your concerns.

Q1. Architectural contribution

What distinguishes ImaGeo from existing joint-attention and per-view/global-attention designs?

Response. We combine per-view processing with bidirectional image–geometry interaction: geometry refines image features, which guide subsequent geometry updates. This simple design supports variable-view inputs using only geometry flow matching. Although related to Hunyuan3D/MM-DiT, to our knowledge we provide the first controlled validation of this design for both single-view 3D generation and multi-view reconstruction, through new feedback ablations and expanded architecture comparisons (Section 4.4, Tables 4–5).

Q2. Geometry-to-image feedback

Does geometry-to-image feedback help when image tokens remain updated?

Response. We interpret this feedback as giving views a shared geometric context for relating their observations to the evolving shape.

The new Stage 1 ablation retains image updates and image-to-geometry attention (Table 5). On 200 GSO/Toys4K objects with four views and a fixed Stage 2 model, disabling feedback worsens all five metrics on both datasets, with pooled CD-L1 58.2% higher. After 10,000 fine-tuning updates, when evaluation metrics had plateaued, it still trails Full on all five metrics, with CD-L1 4.9% higher.

Q3. Comparison with existing architectures

How does ImaGeo compare with Hunyuan/MM-DiT-style designs?

Response. Our new MMDiT-style + view-ID baseline removes per-view image self-attention while retaining shared-QKV joint attention (Table 5). View-ID modulation distinguishes views after per-view processing is removed.

With four views and the same Stage 2 model, the Table 3 ImaGeo checkpoint outperforms this Hunyuan/MM-DiT-inspired variant on all five metrics on both GSO and Toys4K; PSNR improves by 0.479/0.351 dB, respectively.

Q4. Statistical reliability

Are Table 4's gains reliable given the small test set and missing variance estimates?

Response. We reran all three architectures on 20 Objaverse objects and 20 artist-sculpted models with five inference seeds each (600 results), matching inputs, GT coarse geometry, and initial noise. Table 4 reports mean ± SD across five test-set means and paired 95% CIs from 10,000 object-bootstrap resamples after averaging seeds per object.

ImaGeo improves all four metrics over CA 1.5B, with all paired CIs excluding zero: PSNR improves by 1.020 dB [0.945, 1.100] and LPIPS decreases by 18.85%. Against CA 5B, observed quality is comparable and all four CIs include zero.

Q5. Computational cost

Are the gains explained by higher inference compute?

Response. Tables 6–7 report Stage 2 DiT costs on one L20 (FP16, 32,768 geometry tokens). With one view, ImaGeo uses 6.0% more FLOPs and 2.9% more CFG-step time than CA 1.5B, but 46.8% fewer FLOPs and 46.7% less time than CA 5B. Together with Table 4, this shows that the quality gains do not require substantially more inference compute.

Peak allocated memory is 7.756 GiB versus 8.915/19.563 GiB for CA 1.5B/5B. CFG-step latency rises from 7.855 to 15.969 s between one and six views.

Q6. Stage 1, multi-view, and public-data evaluation

Does the architecture evaluation cover Stage 1, multi-view reconstruction, and public data?

Response. Table 5 adds a Stage 1 comparison with the Hunyuan/MM-DiT-inspired variant, improving all five metrics on both GSO and Toys4K (100 objects each). We prioritized this over a naive multi-view CA extension, which we expect to be less effective because fixed, independently encoded features lack explicit cross-view refinement; we did not train that baseline.

Table 4 uses 20 Objaverse objects and 20 artist-sculpted models. The latter add finer details that help distinguish subtle architectural differences.


Thank you again for your consideration. We welcome further questions and feedback.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer z7o6,

We would be grateful if you could share your thoughts on our rebuttal and additional experiments. Please let us know whether any concerns remain, so that we can address them before the discussion closes.

Thank you for your time and consideration.

Best regards, The Authors

I thank the authors for their rebuttal. My questions have been resolved, and I maintain my score.

Comments on the paper

3 comments

Summary of Revisions and Responses

Dear Program Chairs, Area Chair, and Reviewers,

Thank you for your time and constructive feedback. We have revised the manuscript and expanded the experiments to address the concerns shared across the reviews.

1. Clarified contribution and variable-view conditioning

ImaGeo combines per-view image processing with bidirectional image–geometry interaction for single-view 3D generation and multi-view reconstruction. Both denoisers are trained solely with geometry flow matching, without auxiliary camera-prediction or reconstruction losses. Our contribution is this simple, effective conditioning design and its empirical validation.

Experiments with 1, 2, 4, and 6 views show improvements across all five metrics on both GSO and Toys4K as more views are provided (Appendix B.1).

2. Expanded comparisons and statistical evidence

  • We added a 1.5B cross-attention baseline to Table 1, with ImaGeo improving all three normal-map metrics on both datasets.

  • We also reran Table 4 on 20 Objaverse objects and 20 artist-sculpted models, using five inference seeds per architecture and object, and report mean ± SD and paired 95% confidence intervals. ImaGeo improves all four metrics over CA 1.5B, with all four paired intervals excluding zero; PSNR improves by 1.020 dB [0.945, 1.100]. Observed quality is comparable to CA 5B.

  • For multi-view reconstruction, Table 5 adds an MMDiT-style + view-ID variant inspired by Hunyuan/MM-DiT. ImaGeo performs better on all five metrics on both GSO and Toys4K (200 objects in total).

  • We also added Stage 1 geometry-to-image feedback experiments that retain image updates and image-to-geometry attention (Section 4.4, Table 5). On the same 200 objects with four input views and a fixed Stage 2 model, the feedback-disabled variants perform worse than Full across all five metrics on both datasets at the evaluated checkpoints, both before and after 10,000 fine-tuning updates.

3. Reported computational cost

Tables 6–7 report DiT FLOPs, latency, memory, and scaling with view count. In the single-view Stage 2 benchmark, ImaGeo uses only 6.0% more FLOPs and 2.9% more CFG-step time than CA 1.5B, while using 46.8% fewer FLOPs and 46.7% less time than CA 5B. The measured quality gains therefore do not require a substantially larger inference-compute budget.

4. Documented data selection and preprocessing

All architectures in Table 4 use the same training assets and preprocessing, controlling for data curation in this comparison. We have added the data selection, flood-fill reconstruction, and internal-geometry filtering details to Appendix A.4.

Thank you again for considering our revisions. We hope the additional evidence addresses the shared concerns, and welcome further questions and feedback.

Additional Experimental Results

Dear Area Chair and Reviewers,

For convenience, we reproduce the added and updated experimental results below. Table numbers refer to the revised manuscript. CA denotes cross-attention; PSNR, SSIM, and LPIPS use normal maps. Best quality scores and lowest costs are bold within each comparison group.

Table 1 (excerpt): single-view CA baseline. 100 objects per dataset, with pose alignment using GT normal maps.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓
GSOCA 1.5B23.8060.81890.0960
GSOImaGeo24.5180.83760.0855
Toys4KCA 1.5B23.7230.86430.0884
Toys4KImaGeo24.8310.88630.0740

Table 4: Stage 2 architecture comparison. 20 Objaverse and 20 artist-sculpted objects, with five inference seeds per architecture and object (600 outputs). Models share training data and preprocessing, with matched inputs, GT coarse geometry, and initial noise. Entries are mean ± SD across five test-set means; IoU uses silhouettes.

MethodPSNR ↑SSIM ↑LPIPS ↓IoU ↑
CA 1.5B24.6177 ± 0.01430.87118 ± 0.000280.04430 ± 0.000320.97399 ± 0.00017
CA 5B25.6330 ± 0.01390.89408 ± 0.000200.03546 ± 0.000120.97732 ± 0.00006
ImaGeo 1.5B25.6381 ± 0.00730.89370 ± 0.000230.03595 ± 0.000270.97721 ± 0.00006

Paired 95% CIs for ImaGeo minus baseline, using 10,000 object-bootstrap resamples after averaging seeds per object:

BaselineΔPSNRΔSSIMΔLPIPSΔIoU
CA 1.5B[0.9450, 1.0995][0.01967, 0.02576][-0.00975, -0.00702][0.00264, 0.00386]
CA 5B[-0.0527, 0.0779][-0.00142, 0.00071][-0.00028, 0.00119][-0.00052, 0.00035]

Negative LPIPS differences favor ImaGeo. All four metrics improve over CA 1.5B, with CIs excluding zero; all four CIs versus CA 5B include zero.

Table 5: Stage 1 ablations. 100 objects per dataset and four input views; metrics use final meshes from a fixed Stage 2 model. G→I denotes geometry-to-image feedback. Disabling it retains image updates and image-to-geometry attention; FT denotes fine-tuning with feedback disabled.

DatasetVariantCD-L1 ↓CD-L2 ↓PSNR ↑SSIM ↑LPIPS ↓
GSONo first-view alignment0.012600.00032819.6550.78190.1388
GSOMMDiT-style + view-ID0.009340.00019521.3150.80790.1009
GSOw/o G→I0.014410.00053719.2230.77210.1579
GSOw/o G→I + FT0.008870.00016521.5050.81310.0960
GSOImaGeo (full)0.008550.00015521.7940.81640.0923
Toys4KNo first-view alignment0.012610.00063220.3080.80860.1322
Toys4KMMDiT-style + view-ID0.008440.00023621.6740.83750.0903
Toys4Kw/o G→I0.011920.00042319.9170.80460.1330
Toys4Kw/o G→I + FT0.008590.00024921.7650.83700.0903
Toys4KImaGeo (full)0.008090.00021822.0250.84600.0818

Table 6: Stage 2 DiT cost. One view, 32,768 geometry tokens, FP16, one L20. FLOPs are per conditional forward (multiply–add = two operations); latency is per CFG step (batch size two), and memory is peak allocated GPU memory.

MethodTFLOPs ↓CFG step (s) ↓Memory (GiB) ↓
CA 1.5B371.3397.6368.915
CA 5B740.35114.73819.563
ImaGeo 1.5B393.5847.8557.756

Table 7: DiT cost versus view count. Units and measurement conventions follow Table 6; lowest costs are bold within each stage.

StageViewsTFLOPs ↓CFG step (s) ↓Memory (GiB) ↓
1136.7940.8194.034
1242.9880.9474.222
1456.8371.2534.600
1672.6351.6014.977
21393.5847.8557.756
22462.3579.2328.484
24621.92912.3329.963
26810.86915.96911.421

Thank you for considering these additional results; we welcome further feedback.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal and Revised Manuscript

Dear Reviewers,

Thank you for your time and consideration. We would appreciate your feedback on our rebuttal and revised manuscript, particularly whether any concerns remain. We are happy to provide further clarification and look forward to hearing from you.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.