acceptodds
Under review as a conference paper at ICLR 2027

Latent Bridge Trajectory Distillation for 3D Super-Resolution

23 views

Abstract

Synthesizing high-resolution (HR) novel views from given low-resolution (LR) observations with known camera poses is intrinsically ill-posed. Existing 3D Gaussian Splatting (3DGS) super-resolution (SR) methods typically incorporate pretrained restoration priors into the optimization process. However, conventional diffusion-based restoration methods follow a noise-to-HR trajectory that is poorly aligned with the 3DGS optimization initialized from LR observations, and HR details inferred independently across views can be inconsistent. To address these issues, we propose Latent Bridge Trajectory Distillation (LBTD), which progressively distills restoration knowledge along an LR-anchored restoration trajectory into the 3D representation. Specifically, a multi-view latent bridge transports LR observations directly toward the HR distribution, while a cross-view adapter enables pose-nearest views to exchange complementary information for more consistent detail recovery. Trajectory distillation then maintains a persistent bridge trajectory across successive 3DGS optimization stages, where clean HR images are estimated at intermediate steps and updated 3D renderings are fed back to advance the trajectory. For detail enhancement, we propose detail-aware Gaussian splitting, which allocates finer primitives along directions where SR supervision reveals structures beyond the evidence available in the LR observations. Experiments show that LBTD achieves state-of-the-art performance, outperforming 3DSR by 0.31–0.89 dB in PSNR while using fewer Gaussians. In addition, LBTD recovers richer details with fewer distortions, leading to sharper and more faithful HR renderings.

Primary area:
generative models
Keywords:
3D Gaussian Splatting, 3D Super-Resolution, Latent Bridge Matching

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_JNozOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a latent bridge trajectory distillation (LBTD) for 3D Gaussian Splatting super-resolution. It introduces several modules, such as multi-view latent bridge restoration, progressive trajectory distillation, and detail-aware Gaussian splitting, to improve high-resolution rendering quality and cross-view consistency. Experiments demonstrate the effectiveness of the proposed method.

Critical strengths

The LR-anchored restoration trajectory provides an interesting perspective on diffusion-guided 3D super-resolution.

Critical weaknesses
  1. This method relies on large-scale pre-training to generate backbone, but does not fully analyze the total training cost and memory requirements. In ablation studies, the independent improvements from cross-view adapters and detail-aware segmentation are relatively small.
  2. The proposed trajectory distillation lacks a deeper theoretical explanation for its superiority over alternative iterative recovery schemes. The robustness of pseudo HR supervision to hallucinatory structures deserves further research.
Questions for authors

Why does the cross view adapter only provide minor improvements in reconstruction metrics? How does trajectory distillation behave when the recovery model produces inaccurate high-frequency details?

7 comments

Weakness 1

This method relies on large-scale pre training to generate backbone, but does not fully analyze the total training cost and memory requirements. In ablation research, the independent improvement of cross view adapters and detail aware segmentation is relatively small.

Response to Weakness 1

Thank you for raising these concerns. The training of our SR model takes approximately 300 A100 GPU-hours on A100 80GB GPUs. The peak memory usage during training is approximately 66 GiB, with a batch size of 4 images per GPU.

To better examine the improvement of our cross-view adapter and detail-aware splitting, we repeated each configuration with three random seeds on Mip-NeRF360 and Tanks & Temples. We report the mean ± standard deviation:

Mip-NeRF360:

MethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
3DSR26.01 ± 0.080.7381 ± 0.00090.2334 ± 0.00085.108 ± 0.0040.1280 ± 0.000223.58 ± 0.33
+ Single-View SR26.29 ± 0.030.7430 ± 0.00050.2351 ± 0.00025.376 ± 0.0040.1226 ± 0.000322.76 ± 0.29
+ Multi-View SR26.34 ± 0.020.7455 ± 0.00030.2214 ± 0.00025.048 ± 0.0030.1227 ± 0.000221.40 ± 0.21
+ Multi-View SR + TD26.49 ± 0.020.7561 ± 0.00020.1947 ± 0.00024.850 ± 0.0120.1089 ± 0.000320.79 ± 0.22
+ Multi-View SR + TD + Detail-Aware Splitting (ours)26.53 ± 0.010.7570 ± 0.00030.1940 ± 0.00014.885 ± 0.0030.1083 ± 0.000120.05 ± 0.19

Tanks & Temples:

MethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
3DSR22.52 ± 0.040.7521 ± 0.00050.2095 ± 0.00024.128 ± 0.0040.1397 ± 0.000129.09 ± 0.47
+ Single-View SR22.63 ± 0.020.7578 ± 0.00040.2158 ± 0.00024.552 ± 0.0030.1481 ± 0.000126.51 ± 0.40
+ Multi-View SR22.75 ± 0.010.7625 ± 0.00040.2055 ± 0.00024.687 ± 0.0050.1328 ± 0.000124.58 ± 0.19
+ Multi-View SR + TD22.82 ± 0.020.7696 ± 0.00020.1886 ± 0.00013.986 ± 0.0090.1237 ± 0.000222.33 ± 0.13
+ Multi-View SR + TD + Detail-Aware Splitting (ours)22.94 ± 0.010.7710 ± 0.00030.1853 ± 0.00024.015 ± 0.0050.1235 ± 0.000121.56 ± 0.14

To assess the cross-view adapter, we compare Single-View SR with Multi-View SR. To assess detail-aware splitting, we compare Multi-View SR + TD with the full model, where TD refers to trajectory distillation. The cross-view adapter improves PSNR by 0.05 dB on Mip-NeRF360 and 0.12 dB on Tanks & Temples, while reducing FID by 1.36 and 1.93, respectively. Detail-aware splitting improves PSNR by 0.04 dB and 0.12 dB, and reduces FID by 0.74 and 0.77, respectively. Thus, the gains remain positive after averaging over three seeds, with larger PSNR gains on Tanks & Temples. The adapter improves SR supervision by sharing complementary information across views, while detail-aware splitting helps the Gaussian representation capture finer structures. They therefore improve different parts of the pipeline and provide complementary gains in reconstruction quality.

Weakness 2

The proposed trajectory distillation lacks a deeper theoretical explanation for its superiority over alternative iterative recovery schemes. The robustness of pseudo HR supervision to hallucinatory structures deserves further research.

Response to Weakness 2

Our proposed trajectory distillation is different from other iterative recovery schemes, like 3DSR, regarding trajectory initialization and the stage-wise sampling schedule.

  1. Trajectory initialization. Our trajectory starts directly from the encoded LR observation, whereas 3DSR starts from a noise-corrupted latent. Our LR starting state already contains the observed content and coarse structure, providing an anchor for recovering missing details.
  2. Stage-wise sampling schedule. Our TD aligns four SR sampling steps with four GS optimization stages, with two benefits:
    • Sampling efficiency. Each GS stage uses one sampling step to obtain a clean HR endpoint prediction, without completing the entire four-step trajectory. Compared with 3DSR that requires multiple sampling steps (e.g., 2 steps), this reduces SR sampling overhead and achieves higher sampling efficiency.
    • Aligned SR and 3DGS refinement. Each SR sampling step is paired with a corresponding GS optimization stage. The SR prediction guides GS optimization, while the updated 3D rendering guides subsequent bridge-state evolution. By retaining the bridge state across stages, our GS optimization follows the progression of the SR trajectory while also influencing its subsequent evolution, allowing restoration and reconstruction to progressively refine each other.

We discuss the robustness of pseudo-HR supervision to hallucinated structures in our response to the next question.

Question 1

Why does the cross view adapter only provide minor improvements in reconstruction metrics? How does trajectory distillation behave when the recovery model produces inaccurate high-frequency details?

Response to Question 1

Thank you for raising these questions. As shown above, the cross-view adapter provides smaller PSNR gains on Mip-NeRF360 than on Tanks & Temples after GS optimization. Its improvements to SR supervision may become less pronounced after GS optimization. To assess the adapter’s effect before 3D reconstruction, we compare Single-View SR and Multi-View SR directly on LR input views from four datasets, without subsequent 3DGS optimization:

DatasetMethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
Mip-NeRF360Single-View SR29.090.85310.17566.1550.08996.51
Mip-NeRF360Multi-View SR29.340.86670.16835.9920.09006.44
LLFFSingle-View SR29.990.90200.09445.1380.06415.76
LLFFMulti-View SR30.590.90630.08254.9280.05733.72
Tanks & TemplesSingle-View SR25.080.78920.14514.2340.085411.14
Tanks & TemplesMulti-View SR25.460.79390.13804.2460.08048.55
DL3DVSingle-View SR30.850.88500.10965.4970.05764.84
DL3DVMulti-View SR31.260.89290.10025.1810.04713.15

The adapter improves PSNR by 0.25 dB, 0.60 dB, 0.38 dB, and 0.41 dB on Mip-NeRF360, LLFF, Tanks & Temples, and DL3DV, respectively. It also improves SSIM, LPIPS, and FID across all four datasets, demonstrating its benefits across datasets.

Trajectory distillation uses stage-wise supervision and 3D rendering feedback to update SR predictions alongside the shared 3D representation, rather than fitting fixed SR targets. This allows inaccurate details in early SR predictions to be revised as reconstruction proceeds. In Figure 13, trajectory distillation initially produces distorted shapes of the LEGO stud and wheel, but refines these structures in later stages. In contrast, the fixed pseudo-HR target retains its original errors throughout GS optimization. This example illustrates how trajectory distillation can revise inaccurate SR supervision, rather than repeatedly using the same erroneous SR target.

Reviewer_JNozOpenReview ↗

Final Comment

Thank you for your detailed rebuttal and the additional experiments. First, the authors did not provide a revised manuscript incorporating the new experiments, so it is unclear whether they were unaware of the rebuttal requirements. Without an updated version, it is difficult to verify how these issues were addressed in the paper. Second, I cannot find a disclosure statement regarding the use of LLMs or AI in the manuscript, , and Figure 2 appears to rely heavily on AI-generated visual content. I am concerned about whether this complies with the ICLR guidelines and whether it might be rejected.

Regarding the rebuttal, I acknowledge that the additional experiments addressed some of my concerns. However, The improvements were limited, and the computational overhead was significant. The benefits of detail-aware splitting remain relatively modest, and the reported computational analysis did not fully account for the cost of the entire pipeline.

The rationale for trajectory distillation is insufficient. The authors primarily provide intuitive explanations for LR-anchored initialization and stage optimization, but they lack sufficient controlled experiments to demonstrate their respective contributions or to show that they outperform alternative iterative recovery schemes.

Robustness to hallucinated details. Although the qualitative examples in Figure 13 are encouraging, a single example is insufficient to establish the robustness of trajectory distillation to inaccurate or hallucinated structures in pseudo-HR supervision.

Overall, I appreciate the authors’ efforts and acknowledge that some concerns have been partially or substantially addressed. However, due to the lack of a revised draft, the need for clarification regarding key figures, and remaining methodological and robustness issues, I cannot fully endorse the paper’s contributions. Therefore, I maintain my original rating.

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

This paper proposes LBTD (Latent Bridge Trajectory Distillation) for reconstructing high-resolution 3D Gaussian Splatting from low-resolution, multi-view images. The method mainly includes: a latent bridge super-resolution model based on LR constraints, using cross-view attention to enhance multi-view consistency, performing trajectory distillation during 3DGS optimization, and detail-aware Gaussian splitting based on gradient directions. The authors conducted experiments on Mip-NeRF 360, DL3DV, LLFF, and Tanks & Temples, showing certain performance improvements compared to 3DSR, while also reducing the number of Gaussians and speeding up rendering. Ablation studies indicate that the main gains come from the latent bridge SR prior, while the additional benefits from trajectory distillation, cross-view adapter, and detail-aware splitting are relatively limited.

Critical strengths
  1. The overall approach is well-designed and reasonable. It combines LR-constrained generative super-resolution priors with 3DGS optimization, and updates the latent trajectory through 3D rendering feedback, forming a fairly complete closed-loop optimization framework.
  2. The experimental results are competitive. There are consistent improvements over 3DSR across multiple datasets, and the method also has advantages in terms of the number of Gaussians and rendering efficiency.
Critical weaknesses
  1. SOTA comparisons are not sufficient. The paper lacks comparisons with recent relevant 3DGS super-resolution methods, so it currently does not fully support the SOTA claim.
  2. The independent contributions of each component are not very clear. Ablation results show that most of the performance gain comes from the stronger lens-based latent bridge; trajectory distillation, cross-view adapter, and detail-aware splitting only provide minor improvements in PSNR. Therefore, it's unclear whether the main contribution comes from the new optimization framework or the stronger SR prior.
Questions for authors
  1. Were baselines like 3DSR obtained using official results or re-implemented? If re-implemented, please clarify the 3DGS backbone, number of optimization iterations, densification strategy, and SR model configuration to ensure fair comparison.
  2. Can you, under the same lens-based SR backbone, gradually add cross-view attention, trajectory distillation, and detail-aware splitting to more clearly quantify each module's independent contribution?
  3. Why were recent methods like SR3R, IE-SRGS, MVGSR, SRSplat, etc., not compared? If a direct comparison is not possible, please explain the specific reason.
3 comments

Question 1

Were baselines like 3DSR obtained using official results or re-implemented? If re-implemented, please clarify the 3DGS backbone, number of optimization iterations, densification strategy, and SR model configuration to ensure fair comparison.

Response to Question 1

Thank you for asking for clarification. We obtained our baseline results by running the authors’ official GitHub implementations. Specifically, we follow the official training settings and adopt the provided 3DGS backbone with 50K optimization iterations. The densification strategy and SR model configurations are consistent with the official implementation.

Question 2

Can you, under the same lens-based SR backbone, gradually add cross-view attention, trajectory distillation, and detail-aware splitting to more clearly quantify each module's independent contribution?

Response to Question 2 (and Weakness 2)

Thank you for suggesting this ablation. Starting from 3DSR, we first replace the SR model in 3DSR with our single-view Lens-based SR model, then progressively add the cross-view adapter, trajectory distillation, and detail-aware splitting. Across these four variants, we keep the Lens backbone, 3DGS backbone, and number of GS optimization iterations fixed. We repeated per-scene optimization for each configuration using three random seeds on Mip-NeRF360 and Tanks & Temples. The tables report the mean ± standard deviation across seeds:

Mip-NeRF360:

MethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
3DSR26.01 ± 0.080.7381 ± 0.00090.2334 ± 0.00085.108 ± 0.0040.1280 ± 0.000223.58 ± 0.33
+ Single-View SR26.29 ± 0.030.7430 ± 0.00050.2351 ± 0.00025.376 ± 0.0040.1226 ± 0.000322.76 ± 0.29
+ Multi-View SR26.34 ± 0.020.7455 ± 0.00030.2214 ± 0.00025.048 ± 0.0030.1227 ± 0.000221.40 ± 0.21
+ Multi-View SR + TD26.49 ± 0.020.7561 ± 0.00020.1947 ± 0.00024.850 ± 0.0120.1089 ± 0.000320.79 ± 0.22
+ Multi-View SR + TD + Detail-Aware Splitting (ours)26.53 ± 0.010.7570 ± 0.00030.1940 ± 0.00014.885 ± 0.0030.1083 ± 0.000120.05 ± 0.19

Tanks & Temples:

MethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
3DSR22.52 ± 0.040.7521 ± 0.00050.2095 ± 0.00024.128 ± 0.0040.1397 ± 0.000129.09 ± 0.47
+ Single-View SR22.63 ± 0.020.7578 ± 0.00040.2158 ± 0.00024.552 ± 0.0030.1481 ± 0.000126.51 ± 0.40
+ Multi-View SR22.75 ± 0.010.7625 ± 0.00040.2055 ± 0.00024.687 ± 0.0050.1328 ± 0.000124.58 ± 0.19
+ Multi-View SR + TD22.82 ± 0.020.7696 ± 0.00020.1886 ± 0.00013.986 ± 0.0090.1237 ± 0.000222.33 ± 0.13
+ Multi-View SR + TD + Detail-Aware Splitting (ours)22.94 ± 0.010.7710 ± 0.00030.1853 ± 0.00024.015 ± 0.0050.1235 ± 0.000121.56 ± 0.14

TD denotes trajectory distillation. The cross-view adapter improves PSNR by 0.05 dB on Mip-NeRF360 and 0.12 dB on Tanks & Temples, while reducing FID by 1.36 and 1.93, respectively. Adding trajectory distillation increases PSNR by 0.15 dB and 0.07 dB and reduces FID by 0.61 and 2.25, respectively. Detail-aware splitting further improves PSNR by 0.04 dB and 0.12 dB and reduces FID by 0.74 and 0.77, respectively. Together, the three components improve all six metrics over the Single-View SR baseline on both datasets. The adapter shares complementary information across views, trajectory distillation lets SR supervision evolve with the 3D reconstruction, and detail-aware splitting helps represent finer structures more effectively. These improvements therefore go beyond replacing the SR backbone alone.

Question 3

Why were recent methods like SR3R, IE-SRGS, MVGSR, SRSplat, etc., not compared? If a direct comparison is not possible, please explain the specific reason.

Response to Question 3 (and Weakness 1)

Thank you for suggesting these comparisons. We have added comparisons with all four methods on the LLFF dataset:

MethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
SR3R24.300.74230.18235.2410.126531.98
IE-SRGS24.280.75220.18705.0960.122030.63
MVGSR24.580.77460.17815.1930.127527.55
SRSplat24.370.77110.18954.9370.129036.23
LBTD (Ours)24.660.77300.17085.0170.119423.00

Compared with these four methods, our method achieves the best PSNR, LPIPS, DISTS, and FID, and the second-best SSIM and NIQE.

We have re-examined the paper and reconsidered our assessment in light of the authors’ responses and the other reviewers’ comments. The additional ablations using the same Lens backbone and comparisons with recent baselines address our main concerns about the contributions of the proposed components and baseline coverage. Although the gains from individual components remain modest and trajectory distillation shows mixed results across perceptual metrics, we are inclined to raise our rating from 2 (Weak rejection) to 3 (Weak acceptance).

Thank you for your time and valuable feedback. We appreciate your positive assessment of our work and your suggestions on our experiments.

Regarding trajectory distillation, we have added a comparison with 3DSR-style iterative recovery. We replace trajectory distillation in our full model with this recovery scheme, keeping other settings the same. The results are shown below:

DatasetRecovery schemePSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
Mip-NeRF3603DSR-style iterative recovery26.490.75120.22085.4870.117320.84
Mip-NeRF360Trajectory Distillation26.530.75700.19404.8850.108320.05
Tanks & Temples3DSR-style iterative recovery22.710.75200.24424.7200.147424.43
Tanks & TemplesTrajectory Distillation22.940.77100.18534.0150.123521.56
LLFF3DSR-style iterative recovery24.570.76700.21005.6410.134222.89
LLFFTrajectory Distillation24.660.77300.17085.0170.119423.00
DL3DV3DSR-style iterative recovery28.790.87100.16675.8830.114819.01
DL3DVTrajectory Distillation29.110.88400.12245.2070.093311.75

Compared with 3DSR-style iterative recovery, our trajectory distillation improves PSNR, SSIM, LPIPS, NIQE, and DISTS on all four datasets. These results support the benefit of trajectory distillation over 3DSR-style iterative recovery.

We have also added more visual examples from all four datasets in Figures 13–16 (pp. 35–37) of the revised manuscript. These examples show that trajectory distillation initially produces inaccurate details but refines them in later stages. In contrast, fixed pseudo-HR targets retain their original errors.

Reviewer_4FKQOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Latent Bridge Trajectory Distillation (LBTD) for 3D super-resolution. It combines an LR-to-HR latent bridge, cross-view attention, persistent trajectory distillation with rendering feedback, and gradient-guided Gaussian splitting. The overall results are competitive, but the evidence establishing the distinct contributions of the proposed mechanisms remains insufficient.

Critical strengths
  • Clear motivation and coherent design. Coupling restoration with an evolving 3D representation is a reasonable approach to addressing content drift and cross-view inconsistency.
  • Competitive overall performance. Across four datasets, the method improves PSNR over 3DSR by approximately 0.31–0.89 dB while using 13.6%–27.1% fewer Gaussians.
  • Useful supplementary evaluation. The appendix includes trajectory and initialization ablations, multi-view consistency measurements, and rendering-speed comparisons.
Critical weaknesses
  • Limited methodological novelty and incomplete attribution. Latent bridge matching is established in LBM, while alternating restoration and reconstruction with rendering feedback appears in 3DSR. Matched-backbone and matched-training comparisons are needed to separate the proposed mechanisms from the benefits of the Lens teacher and additional training.
  • Small component-level gains without uncertainty estimates. In Table 2, the cross-view adapter and detail-aware splitting improve PSNR by only 0.01 dB and 0.02 dB, respectively. Some perceptual metrics improve, but repeated-run statistics are necessary to establish robustness.
  • Mixed evidence for trajectory distillation. In Table 9, the full method improves PSNR by 0.12 dB over fixed pseudo-HR supervision, but worsens LPIPS, DISTS, and NIQE. This indicates a trade-off rather than a consistent improvement in detail quality.
  • Heuristic interpretation and incomplete cost analysis. Orthogonality to the LR-loss gradient does not establish that the remaining component corresponds to missing high-frequency detail. The paper also lacks a complete comparison of training, restoration, and reconstruction costs.
Questions for authors
  1. Can you compare bridge-based and noise-initialized restoration using the same Lens backbone, training data, and optimization budget to isolate the contribution of the bridge formulation?
  2. Are the small gains from cross-view attention and detail-aware splitting reproducible across random seeds and datasets? Please report uncertainty estimates.
  3. How do you explain the worse LPIPS, DISTS, and NIQE relative to fixed pseudo-HR supervision in Table 9? Under what conditions is trajectory distillation preferable?
  4. What evidence supports interpreting the projected gradient as “detail-specific”? How does it compare with the raw pseudo-HR gradient and deterministic longest-axis splitting under a matched Gaussian budget?
  5. What are the end-to-end runtime and peak memory requirements, including teacher adaptation, multi-view SR inference, and per-scene optimization, compared with the strongest baselines?
6 comments

Question 1

Can you compare bridge-based and noise-initialized restoration using the same Lens backbone, training data, and optimization budget to isolate the contribution of the bridge formulation?

Response to Question 1 (and Weakness 1)

Thank you for suggesting this comparison. We compare bridge-based and noise-initialized restoration using the same pretrained Lens backbone, training data, and optimization budget. The bridge-based variant starts from the encoded LR image, whereas the noise-initialized variant starts from Gaussian noise and receives the LR latent as conditioning through an additional trainable input projection.

We evaluate direct super-resolution on LR views from Mip-NeRF360, without 3DGS optimization:

MethodPSNR ↑SSIM ↑LPIPS ↓NIQE ↓DISTS ↓FID ↓
Noise-initialized27.070.74110.25076.4720.181815.08
Bridge-based29.090.85310.17566.1550.08996.51

The bridge-based variant performs better on all six metrics. These results support the benefit of using the LR observation as the starting point for restoration. Because the LR image already contains the scene content and coarse structure, restoration can build on this information and focus on recovering missing details.

Question 2

Are the small gains from cross-view attention and detail-aware splitting reproducible across random seeds and datasets? Please report uncertainty estimates.

Response to Question 2 (and Weakness 2)

Thank you for suggesting this evaluation. Yes, we see consistent gains across different seeds and datasets. We repeated each configuration with three random seeds on Mip-NeRF360 and Tanks & Temples. We report the mean ± standard deviation:

Mip-NeRF360:

MethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
3DSR26.01 ± 0.080.7381 ± 0.00090.2334 ± 0.00085.108 ± 0.0040.1280 ± 0.000223.58 ± 0.33
+ Single-View SR26.29 ± 0.030.7430 ± 0.00050.2351 ± 0.00025.376 ± 0.0040.1226 ± 0.000322.76 ± 0.29
+ Multi-View SR26.34 ± 0.020.7455 ± 0.00030.2214 ± 0.00025.048 ± 0.0030.1227 ± 0.000221.40 ± 0.21
+ Multi-View SR + TD26.49 ± 0.020.7561 ± 0.00020.1947 ± 0.00024.850 ± 0.0120.1089 ± 0.000320.79 ± 0.22
+ Multi-View SR + TD + Detail-Aware Splitting (ours)26.53 ± 0.010.7570 ± 0.00030.1940 ± 0.00014.885 ± 0.0030.1083 ± 0.000120.05 ± 0.19

Tanks & Temples:

MethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
3DSR22.52 ± 0.040.7521 ± 0.00050.2095 ± 0.00024.128 ± 0.0040.1397 ± 0.000129.09 ± 0.47
+ Single-View SR22.63 ± 0.020.7578 ± 0.00040.2158 ± 0.00024.552 ± 0.0030.1481 ± 0.000126.51 ± 0.40
+ Multi-View SR22.75 ± 0.010.7625 ± 0.00040.2055 ± 0.00024.687 ± 0.0050.1328 ± 0.000124.58 ± 0.19
+ Multi-View SR + TD22.82 ± 0.020.7696 ± 0.00020.1886 ± 0.00013.986 ± 0.0090.1237 ± 0.000222.33 ± 0.13
+ Multi-View SR + TD + Detail-Aware Splitting (ours)22.94 ± 0.010.7710 ± 0.00030.1853 ± 0.00024.015 ± 0.0050.1235 ± 0.000121.56 ± 0.14

To assess the cross-view adapter, we compare Single-View SR with Multi-View SR. To assess detail-aware splitting, we compare Multi-View SR + TD with the full model, where TD refers to trajectory distillation. The cross-view adapter improves PSNR by 0.05 dB on Mip-NeRF360 and 0.12 dB on Tanks & Temples, while reducing FID by 1.36 and 1.93, respectively. Detail-aware splitting improves PSNR by 0.04 dB and 0.12 dB, and reduces FID by 0.74 and 0.77, respectively. Thus, the gains remain positive after averaging over three seeds, with larger PSNR gains on Tanks & Temples.

We additionally evaluate the cross-view adapter through direct SR on LR views from 4 datasets, without subsequent 3DGS optimization:

DatasetMethodPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
Mip-NeRF360Single-View SR29.090.85310.17566.1550.08996.51
Mip-NeRF360Multi-View SR29.340.86670.16835.9920.09006.44
LLFFSingle-View SR29.990.90200.09445.1380.06415.76
LLFFMulti-View SR30.590.90630.08254.9280.05733.72
Tanks & TemplesSingle-View SR25.080.78920.14514.2340.085411.14
Tanks & TemplesMulti-View SR25.460.79390.13804.2460.08048.55
DL3DVSingle-View SR30.850.88500.10965.4970.05764.84
DL3DVMulti-View SR31.260.89290.10025.1810.04713.15

The adapter improves PSNR by 0.25 dB, 0.60 dB, 0.38 dB, and 0.41 dB on Mip-NeRF360, LLFF, Tanks & Temples, and DL3DV, respectively. These results provide additional evidence that cross-view interaction benefits restoration across datasets, because nearby views can provide complementary information that helps resolve details that are ambiguous in a single LR image.

Question 3

How do you explain the worse LPIPS, DISTS, and NIQE relative to fixed pseudo-HR supervision in Table 9? Under what conditions is trajectory distillation preferable?

Response to Question 3 (and Weakness 3)

Thank you for raising this question. In Table 9, trajectory distillation improves PSNR by 0.12 dB, and reduces FID by 0.31, while LPIPS, DISTS, and NIQE increase by 0.0029, 0.0032, and 0.091, respectively. Rendering feedback constrains subsequent SR predictions with the evolving 3D representation and can also constrain fine texture. A fixed pseudo-HR target may contain visually sharp but inaccurate details. Because it is generated only once, these errors remain unchanged in the supervision and cannot be revised through 3D rendering feedback. In contrast, trajectory distillation updates its SR predictions through stage-wise supervision and 3D rendering feedback, allowing early inaccuracies to be revised rather than remaining fixed.

Trajectory distillation is preferable for faithful reconstruction, and is particularly advantageous when an initial SR prediction contains inaccurate structural details. In Figure 13, trajectory distillation initially produces distorted shapes of the LEGO stud and wheel, but refines these structures in later stages, whereas the errors in the fixed pseudo-HR target remain unchanged.

Here we compare the fixed pseudo-HR target and the trajectory distillation predictions at each stage during per-scene optimization, using the kitchen scene in Mip-NeRF360:

Setting (a): SR predictions evaluated against HR references

VariantPSNR↑SSIM↑LPIPS↓NIQE↓DISTS↓FID↓
Fixed pseudo-HR27.590.78880.11475.0090.07077.05
TD Stage 127.430.80230.11515.8710.08306.67
TD Stage 228.910.83160.11115.6180.07686.02
TD Stage 330.220.86220.10815.2360.07675.36
TD Stage 430.540.87470.10705.1120.07194.95

Setting (b): LR observation consistency after downsampling the SR predictions

VariantPSNR↑SSIM↑LPIPS↓
Fixed pseudo-HR34.990.97840.0106
TD Stage 135.510.98300.0086
TD Stage 235.960.98670.0079
TD Stage 336.080.98730.0060
TD Stage 436.120.98680.0065

Setting (a) evaluates the accuracy of the SR predictions against the HR references. TD Stage 1 refers to the SR predictions for trajectory distillation stage 1. In Setting (b), we downsample the SR outputs and compare them with the input LR images to check whether the original content is preserved. In Setting (a), trajectory distillation revises the SR predictions across stages. Compared with fixed pseudo-HR, the final stage improves PSNR from 27.59 dB to 30.54 dB and SSIM from 0.7888 to 0.8747, while reducing LPIPS from 0.1147 to 0.1070 and FID from 7.05 to 4.95. In Setting (b), the final trajectory distillation stage also improves PSNR from 34.99 dB to 36.12 dB and SSIM from 0.9784 to 0.9868, while reducing LPIPS from 0.0106 to 0.0065. These results show that trajectory distillation produces more faithful SR predictions and better preserves the input content. This is because the shared 3D reconstruction is constrained by the input LR images. Rendering feedback from this reconstruction helps preserve the observed content and revise inaccurate details in later SR predictions.

Comments on the paper

1 comment

Dear Reviewers,

The authors have submitted their responses to your reviews. Please take some time to carefully examine their rebuttals and engage in the discussion by sharing your thoughts, remaining concerns, or follow-up questions.

As the discussion deadline is approaching, I would appreciate your timely participation to allow sufficient time for further exchanges with the authors.

Thank you for your time and valuable contributions.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.