RLGS: Reinforcement Learning-Based Adaptive Hyperparameter Tuning for Gaussian Splatting

Changyang Li; Huangying Zhan; Qingan Yan; Yi Xu; Zhan Li

REVIEW 4 major objections 3 minor 12 references

RLGS: Reinforcement Learning-Based Adaptive Hyperparameter Tuning for Gaussian Splatting

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RLGS argues that hyperparameter tuning for 3D Gaussian Splatting can be cast as an online reinforcement learning problem, where lightweight policy modules dynamically scale learning rates and densification thresholds during training, improv

desk verdict RLGS is a modest but credible engineering contribution: it transfers existing policy-gradient hyperparameter control to 3DGS, reports small consistent gains across backbones/datasets, and deserves a careful referee despite nontrivial reproducibility gaps and an unvalidated reward proxy. read the letter →

arxiv 2508.04078 v1 pith:RRQEXFBC submitted 2025-08-06 cs.GR cs.CVcs.LG

Zhan Li , Huangying Zhan , Changyang Li , Qingan Yan , Yi Xu This is my paper

classification cs.GRcs.CVcs.LG

keywords reinforcementlearninghyperparametertuning3DGaussianSplattingpolicygradientadaptiveratedensificationthresholdnovelviewsynthesisrenderingquality

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

RLGS argues that hyperparameter tuning for 3D Gaussian Splatting should be treated as an online decision problem rather than an offline search. It introduces two plug-and-play policy modules—RLLR, which scales the learning rate, and RLDS, which scales densification thresholds—that observe the previous training loss and current iteration, sample scale factors, and receive a reward equal to the rendering-error improvement over default hyperparameters measured on withheld views after 20 simulated training steps. The paper reports consistent gains: 0.7 dB PSNR over Taming-3DGS on Tanks and Temples, 0.21 dB on the 140-scene DL3DV benchmark, and 0.23 dB PSNR with 24% lower LPIPS over 3DGS-MCMC on Mip-NeRF360. If correct, this makes per-scene hyperparameter tuning automatic and model-agnostic, improving quality without changing the Gaussian count or the underlying architecture.

What carries the argument

The load-bearing machinery is the MDP formulation of the 3DGS training loop, realized by two lightweight policy modules. RLLR and RLDS share a design: a GRU encoder takes the state (previous phase's loss, current iteration) and produces residual adjustments to learnable base means and log-variances; actions are sampled from the resulting Gaussian and applied as multiplicative scaling factors to the original hyperparameters. The second key element is the reward: after each sampled action, K=20 steps of 3DGS training are simulated and the improvement over default hyperparameters is measured on two reward views held out from the training set and reshuffled every Ishuffle=1000 steps. This makes

What would settle it

On a fixed TNT scene, sample many hyperparameter perturbations, measure each one's PSNR change on the two reward views after 20 steps, and also measure final test PSNR after full training; if the rank correlation between the short-horizon proxy and the final result is weak or negative, the policy's reward does not carry the signal needed for the claimed improvements to generalize.

Watch

Extended reading notes

Core claim

The paper's central claim is that a reinforcement-learning controller can select 3DGS hyperparameters on the fly and beat both default settings and standard search methods. The optimization of a 3DGS scene is posed as a Markov decision process where, every K=20 training steps, a lightweight policy network adjusts five learning-rate parameters (position, scaling, rotation, opacity, spherical-harmonic base coefficients) or two densification parameters (density threshold and external scaling factor). Policies are trained with REINFORCE plus an entropy bonus, with the reward defined as the improvement in rendering error on a pair of withheld reward views relative to the default hyperparameters;

Load-bearing premise

The reward signal computed after only 20 simulated training steps on two reward views (reshuffled every 1000 steps) faithfully predicts which hyperparameter changes will improve final rendering on held-out test views.

Editorial extensions

If this is right

RLGS is a drop-in module for existing 3DGS variants; it improves both Taming-3DGS and 3DGS-MCMC without architectural changes or Gaussian-count changes.
It outperforms random search and Bayesian optimization while using a single trial instead of 64, making per-scene tuning practical.
It yields gains in scenes where increasing Gaussian count has saturated, decoupling quality from capacity.
The modular policy design can be pointed at other continuous hyperparameters, as shown when the densification head is swapped for a regularization-term tuner on 3DGS-MCMC.
It reduces reliance on expert, scene-specific hyperparameter selection, a stated barrier to broader 3DGS adoption.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The K=20/two-view reward is a proxy for final quality; varying K and the reward-view count, and correlating proxy reward with final test PSNR across random hyperparameter perturbations, would show how much of the gain depends on this proxy being accurate.
Because reward views are periodically reshuffled, the method resembles continual cross-validation inside training; the same online controller could be tried on other differentiable rendering pipelines where short simulated rollouts are cheap.
The paper acknowledges the policy is a black box; a direct comparison against a hand-specified controller that adjusts learning rate from loss trends would isolate how much of the gain comes from learning versus from adaptivity itself.
The fact that gains persist at saturated Gaussian counts hints the policy reshapes optimization dynamics; testing whether a policy trained on one scene transfers to unseen scenes without retraining would reveal whether it captures scene-agnostic rules.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

I read this with the expectation of either a hype-driven claim or a trivial wrapper. It's neither. The core idea is a legitimate new application: use two lightweight policy-gradient agents to scale learning rates and densification thresholds during 3DGS training, rather than tuning them by hand or via full-search HPO. The related work correctly credits Daniel et al. and Wu et al. for the RL-for-LR-control lineage, so the novelty claim is appropriately scoped to "first RL-based HPO for 3DGS." The two-module design (RLLR and RLDS) is simple and sensible, and the ablations in Table 4 actually support the design choices: each removed component costs something, and the effect sizes are plausible.

The empirical results are real, as far as they go: +0.7 dB on TNT, +0.21 dB on DL3DV, +0.23 dB on MipNeRF360 with 3DGS-MCMC, all under fixed Gaussian budgets. The comparison with random search and Bayesian optimization (64 trials each) is fair and actually favorable to the method, since RLGS uses only one training run. That's a useful result for practitioners.

Now the soft spots, in proportion. The biggest one is the reward proxy. The policy is trained on a signal computed after K=20 simulated training steps on just two reward views, and this signal is the sole driver of action selection. The paper never checks whether this short-horizon, two-view reward correlates with final test-view PSNR. That's a real gap. It's possible the gains come from the proxy accidentally selecting reasonable hyperparameters, but it's also possible the mechanism is mostly noise. The stress-test note is right about this. However, it's not a fatal flaw because the reported improvements are measured at the end of full training on held-out views, so the empirical claim stands regardless of proxy validity. What is unsupported is the attribution that the RL objective is what makes it work.

Reproducibility is below the bar for this field. No code, no error bars, no per-scene baseline numbers in the main text (the supplementary has per-scene ours-only, not baselines), and no compute-cost comparison against the 64-trial baselines. The simulated rollouts must be expensive, and the paper doesn't report wall-clock or FLOP overhead. Also, the inner-loop lengths Nlr and NDS are never specified, and the rollback/copy mechanism for evaluating multiple candidate actions from the same model state is not described. These are addressable, but they need to be fixed before anyone should build on this.

Overall, the central empirical claim is plausible and the paper is honest about its scope. Who is it for? 3DGS researchers and systems people who want an automated alternative to hand-tuning; also HPO folks interested in online control. I'd send it to peer review, with a request for code, error bars, compute cost, and a direct validation of the reward proxy. The claims are not groundbreaking, but they are useful and the paper is clear-headed about its contribution.

Recommendation: accept as a conditional, after requested experiments and code release. I would not desk-reject it.

Referee Report

4 major / 3 minor

Summary. The paper proposes RLGS, a reinforcement-learning framework that adaptively tunes hyperparameters during 3D Gaussian Splatting (3DGS) training. Two lightweight policy modules (RLLR and RLDS) observe the previous phase's loss and the current iteration, sample multiplicative scaling factors for learning-rate and densification hyperparameters, and are trained with policy gradients using a reward defined as reconstruction improvement after K=20 simulated training steps on two withheld reward views. The method is plugged into Taming-3DGS and 3DGS-MCMC without architectural changes and evaluated on TNT, Deep Blending, DL3DV-140, and Mip-NeRF360. The central claim is that RLGS consistently improves rendering quality, e.g., +0.7 dB PSNR on the TNT dataset with Taming-3DGS and +0.21 dB on DL3DV-140, while also outperforming random search and Bayesian optimization baselines.

Significance. If the claimed results hold, RLGS addresses a real and practical gap: 3DGS hyperparameters are scene-sensitive and are still manually tuned in most pipelines. The paper's strengths include a clean plug-and-play design, an ablation study isolating each policy module and design choice, evaluation across two 3DGS backbones, and a large-scale dataset (DL3DV-140). The quantitative gains, especially the +0.7 dB PSNR on TNT, are potentially useful for practitioners. The approach is not derivational circular: the improvement claim is empirical. However, the evidence is incomplete in ways that bear directly on the claim, and the core reward signal is not validated as a faithful predictor of final rendering quality.

major comments (4)

[Section 3.2 / Section 4.6] The reward is computed after only K=20 simulated training steps on a reward set of size 2, yet it is the sole signal driving policy updates and the per-phase selection of (hlr, hds). The paper provides no evidence that this short-horizon, two-view proxy correlates with final test-view PSNR after the full 30,000-step training run. With K=20 covering 0.067% of training and reward views reshuffled only every Ishuffle=1000 steps, the reward may reflect transient loss dips or overfitting to two views rather than genuine long-horizon improvement. This is load-bearing: without this validation the claimed gains cannot be attributed to the RL mechanism. Please report the correlation between the proxy reward and final held-out PSNR (e.g., across phases or scenes), or provide an ablation varying K and reward-set length.
[Section 3.1 / Section 4.6] The inner loop lengths Nlr and NDS, which determine how many candidate actions are simulated per policy phase, are never specified. These parameters control the compute budget of RLGS and directly affect the quality of the selected hyperparameters. Without them the method is not reproducible and the reported total training time (25 minutes for the Train scene) cannot be interpreted. Please report Nlr, NDS, and the total number of simulated K-step evaluations per scene, together with a compute comparison against the RS (64 trials) and BO (64 trials) baselines.
[Section 4 / Table 1 / Table 5] The main text reports only averaged results. The per-scene tables in the supplement (Tables 5 and 6) list only RLGS-enhanced numbers, not the baseline TMGS numbers, so the reader cannot see whether the improvement is consistent across scenes or dominated by a single scene. Please include per-scene baseline PSNR/SSIM/LPIPS and the per-scene deltas, or make the supplement tables include both. This is needed to support the claim of 'consistent improvements.'
[Section 3.2 / Section 3.3] To evaluate several candidate actions from the same model state, the optimizer must either roll back to the state before each action or maintain multiple copies of the 3DGS model. The paper does not state how this is implemented. If the reward for one action is measured while the model has already advanced from a previous candidate's simulation, the estimated improvement is contaminated and the policy-gradient update is not correctly attributed. Please specify the rollback/copy mechanism, or, if it is not used, explain how the reward M(h) - M(horig) is computed in isolation for each action.

minor comments (3)

[Table 3] On the MipNeRF360 experiment, 3DGS-MCMC + Ours improves PSNR and LPIPS but SSIM drops from 0.900 to 0.897. The text presents the result as a clear improvement without mentioning this metric decrease; please discuss it.
[Figure 3] The labels '(a) Initial training views', '(b) Split...', '(c) After...' are positioned beneath the figure and the sub-captions are easy to confuse with the text in the figure. Please clarify the figure layout.
[Section 4.6] Minor typos: 'Ishuffle to 1000' is awkward; and in Section 3.1 the reward notation RLR j,t and RDS j,t is defined in prose but the subscripts are not used consistently in Eq. (4).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are empirical results on held-out test views, not quantities forced by the reward definition.

full rationale

The paper's central claim is that RLGS improves 3DGS rendering quality, e.g., +0.7 dB PSNR on TNT with Taming-3DGS. The derivation chain is empirical rather than formal. The reward in Eq. 4 is defined as the improvement over default hyperparameters on reward views, and the policy is optimized to maximize that reward. However, the reported metrics (PSNR, SSIM, LPIPS) are computed on held-out test views (every 8th image) after the full 30,000-step training run, not on the reward views. Reward views are a subset of training views, reshuffled every Ishuffle=1000 steps, and are distinct from test views. Thus the final evaluation is not equal to the optimized reward by construction. No fitted parameter is renamed as a prediction: the policy is trained online per scene and evaluated on that scene's held-out views, which is standard protocol and not circular. The self-citations in the reference list (Hou et al. 2024; Lin et al. 2024) appear only in related-work enumeration and are not load-bearing for the method or results. The skeptical concern about K=20 simulated steps and two reward views is a question of whether the proxy reward transfers to final rendering quality; that is an empirical validity concern, not a circularity concern. A noisy or unrepresentative reward could weaken the method, but it does not make the reported improvement identical to the input by definition. Therefore, no circular step can be exhibited, and the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends on hand-chosen policy hyperparameters and the K-step reward proxy. No new physical entities are introduced.

free parameters (6)

Policy phase length K = 20 training steps
Hand-chosen; larger values reduce policy updates, smaller increase compute; affects reward relevance.
Reward view reshuffle interval Ishuffle = 1000 steps
Empirically set; smaller values destabilize the reward signal (Sec 4.6).
Reward set length = 2 views
Set to 2; small enough to keep reward computation cheap but may be noisy.
Policy network learning rate = 1e-4
Fixed; not studied.
Gradient clipping norm = 2.4
Chosen to stabilize policy training; no sensitivity analysis.
Inner loop lengths Nlr, NDS = not specified
Maximum lengths mentioned in Sec 3.1 but no values given in main text or supplementary; central to compute cost.

assumptions (4)

ad hoc to paper K-step simulated training is a faithful proxy for final 3DGS quality
The policy selects actions using reward after 20 training steps; if this proxy is inaccurate, chosen hyperparameters may not produce final gains. Introduced in Sec 3.2.
domain assumption Reward views sampled from training views provide a valid online signal
The method withholds and reshuffles a small set of training views for reward computation, assuming this cross-validation-style signal estimates test performance without biasing training. Sec 3.2.
domain assumption 3DGS training can be modeled as an MDP with state (previous loss, iteration)
Policy gradient formulation assumes the state is Markovian; loss and iteration may not capture full training dynamics. Sec 3.1.
standard math Standard 3DGS equations are taken as given
Appendix B restates Kerbl et al. 2023 equations; no new derivation needed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RLGS: Reinforcement Learning-Based Adaptive Hyperparameter Tuning for Gaussian Splatting." pith.science (2026). https://pith.science/paper/RRQEXFBC

@misc{pith2026250804078,
  author       = {Pith},
  title        = {Pith review of: RLGS: Reinforcement Learning-Based Adaptive Hyperparameter Tuning for Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RRQEXFBC}},
  note         = {Machine review of arXiv:2508.04078}
}

read the original abstract

Hyperparameter tuning in 3D Gaussian Splatting (3DGS) is a labor-intensive and expert-driven process, often resulting in inconsistent reconstructions and suboptimal results. We propose RLGS, a plug-and-play reinforcement learning framework for adaptive hyperparameter tuning in 3DGS through lightweight policy modules, dynamically adjusting critical hyperparameters such as learning rates and densification thresholds. The framework is model-agnostic and seamlessly integrates into existing 3DGS pipelines without architectural modifications. We demonstrate its generalization ability across multiple state-of-the-art 3DGS variants, including Taming-3DGS and 3DGS-MCMC, and validate its robustness across diverse datasets. RLGS consistently enhances rendering quality. For example, it improves Taming-3DGS by 0.7dB PSNR on the Tanks and Temple (TNT) dataset, under a fixed Gaussian budget, and continues to yield gains even when baseline performance saturates. Our results suggest that RLGS provides an effective and general solution for automating hyperparameter tuning in 3DGS training, bridging a gap in applying reinforcement learning to 3DGS.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

12 extracted references · 6 canonical work pages

[5]

ACM Trans

Scalable Inside-out Image-Based Rendering. ACM Trans. Graph., 35(6). Heigl, B.; Koch, R.; Pollefeys, M.; Denzler, J.; and Gool, L. J. V . 1999. Plenoptic Modeling and Rendering from Image Sequences Taken by Hand-Held Camera. In Mustererken- nung 1999, 21. DAGM-Symposium, 94–101. Berlin, Heidel- berg: Springer-Verlag. ISBN 3540663819. H¨ollein, L.; Bo ˇziˇ...

work page 1999
[6]

arXiv preprint arXiv:2409.12892

3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiv preprint arXiv:2409.12892. Hou, Q.; Rauwendaal, R.; Li, Z.; Le, H.; Farhadzadeh, F.; Porikli, F.; Bourd, A.; and Said, A. 2024. Sort-free Gaus- sian Splatting via Weighted Sum Rendering. arXiv preprint arXiv:2410.18931. Huang, B.; Yu, Z.; Chen, A.; Geiger, A.; and Gao, S. 2024...

arXiv 2024
[8]

Advances in Neural Information Processing Systems , 37: 80965–80986

3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems , 37: 80965–80986. Kheradmand, S.; Vicini, D.; Kopanas, G.; Lagun, D.; Yi, K. M.; Matthews, M.; and Tagliasacchi, A. 2025. Stochastic- Splats: Stochastic Rasterization for Sorting-Free 3D Gaus- sian Splatting. arXiv preprint arXiv:2503.24366. Knapitsch, A....

arXiv 2025
[11]

Scene PSNR SSIM LPIPS Count (M) drjohnson 29.88 0.910 0.232 3.27 playroom 30.65 0.913 0.234 2.33 Table 6: Per-scene quantitative results on the Deep Blending Dataset (Hedman et al

enhanced with our method. Scene PSNR SSIM LPIPS Count (M) drjohnson 29.88 0.910 0.232 3.27 playroom 30.65 0.913 0.234 2.33 Table 6: Per-scene quantitative results on the Deep Blending Dataset (Hedman et al. 2018) using TMGS (Mallick et al

work page 2018
[12]

enhanced with our method. Scene PSNR SSIM LPIPS Count (M) bicycle 26.30 0.810 0.163 5.90 garden 28.36 0.887 0.087 5.20 stump 27.60 0.818 0.166 4.75 room 32.99 0.940 0.168 1.50 counter 29.70 0.927 0.159 1.20 kitchen 32.64 0.941 0.105 1.80 bonsai 33.26 0.955 0.160 1.30 Table 7: Per-scene quantitative results on the MipNeRF 360 (Barron et al. 2022) dataset u...

work page 2022
[1996]

In Siggraph, volume 96, 43–54

The lumigraph. In Siggraph, volume 96, 43–54. Graves, A.; Bellemare, M. G.; Menick, J.; Munos, R.; and Kavukcuoglu, K. 2017. Automated curriculum learning for neural networks. arXiv preprint arXiv:1704.03003. Hedman, P.; Philip, J.; Price, T.; Frahm, J.-M.; Drettakis, G.; and Brostow, G. 2018. Deep blending for free-viewpoint image-based rendering. In SIG...

arXiv 2017
[2016]

Advances in neural information processing systems , 29

Learning to learn by gradient descent by gradient de- scent. Advances in neural information processing systems , 29. Barron, J. T.; Mildenhall, B.; Verbin, D.; Srinivasan, P. P.; and Hedman, P. 2022. Mip-nerf 360: Unbounded anti- aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5470–5479...

arXiv 2022
[2018]

In Proceedings of the International Conference on Machine Learning (ICML)

Efficient Neural Architecture Search via Parameter Sharing. In Proceedings of the International Conference on Machine Learning (ICML). Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal Policy Optimization Algorithms. arXiv preprint arXiv:1707.06347. Sitzmann, V .; Thies, J.; Heide, F.; Nießner, M.; Wetzstein, G.; and Zollh...

arXiv 2017

Show all 12 references

[2019]

In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2623–2631

Optuna: A next-generation hyperparameter optimiza- 7 tion framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2623–2631. Andrychowicz, M.; Denil, M.; Gomez, S.; Hoffman, M. W.; Pfau, D.; Schaul, T.; Shillingford, B.; ...
[2023]

ACM Transactions on Graphics, 42(4)

3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4). Kerbl, B.; Meuleman, A.; Kopanas, G.; Wimmer, M.; Lan- vin, A.; and Drettakis, G. 2024. A hierarchical 3d gaussian representation for real-time rendering of very large datasets. ...

2024
[2024]

Advances in neural infor- mation processing systems, 37: 140138–140158

Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and 200+ fps. Advances in neural infor- mation processing systems, 37: 140138–140158. Flynn, J.; Broxton, M.; Debevec, P.; DuVall, M.; Fyffe, G.; Overbeck, R.; Snavely, N.; and Tucker, R. 2019. Deepview: View ...

2019
[2025]

arXiv:2501.13975

3DGS 2: Near Second-order Converging 3D Gaussian Splatting. arXiv:2501.13975. Levoy, M.; and Hanrahan, P. 1996. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, 31–42. ACM. Li, L.; Jamieson, K.; DeSalvo, G.; R...

1996 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [5]

ACM Trans

Scalable Inside-out Image-Based Rendering. ACM Trans. Graph., 35(6). Heigl, B.; Koch, R.; Pollefeys, M.; Denzler, J.; and Gool, L. J. V . 1999. Plenoptic Modeling and Rendering from Image Sequences Taken by Hand-Held Camera. In Mustererken- nung 1999, 21. DAGM-Symposium, 94–101. Berlin, Heidel- berg: Springer-Verlag. ISBN 3540663819. H¨ollein, L.; Bo ˇziˇ...

work page 1999

[2] [6]

arXiv preprint arXiv:2409.12892

3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiv preprint arXiv:2409.12892. Hou, Q.; Rauwendaal, R.; Li, Z.; Le, H.; Farhadzadeh, F.; Porikli, F.; Bourd, A.; and Said, A. 2024. Sort-free Gaus- sian Splatting via Weighted Sum Rendering. arXiv preprint arXiv:2410.18931. Huang, B.; Yu, Z.; Chen, A.; Geiger, A.; and Gao, S. 2024...

arXiv 2024

[3] [8]

Advances in Neural Information Processing Systems , 37: 80965–80986

3d gaussian splatting as markov chain monte carlo. Advances in Neural Information Processing Systems , 37: 80965–80986. Kheradmand, S.; Vicini, D.; Kopanas, G.; Lagun, D.; Yi, K. M.; Matthews, M.; and Tagliasacchi, A. 2025. Stochastic- Splats: Stochastic Rasterization for Sorting-Free 3D Gaus- sian Splatting. arXiv preprint arXiv:2503.24366. Knapitsch, A....

arXiv 2025

[4] [11]

Scene PSNR SSIM LPIPS Count (M) drjohnson 29.88 0.910 0.232 3.27 playroom 30.65 0.913 0.234 2.33 Table 6: Per-scene quantitative results on the Deep Blending Dataset (Hedman et al

enhanced with our method. Scene PSNR SSIM LPIPS Count (M) drjohnson 29.88 0.910 0.232 3.27 playroom 30.65 0.913 0.234 2.33 Table 6: Per-scene quantitative results on the Deep Blending Dataset (Hedman et al. 2018) using TMGS (Mallick et al

work page 2018

[5] [12]

enhanced with our method. Scene PSNR SSIM LPIPS Count (M) bicycle 26.30 0.810 0.163 5.90 garden 28.36 0.887 0.087 5.20 stump 27.60 0.818 0.166 4.75 room 32.99 0.940 0.168 1.50 counter 29.70 0.927 0.159 1.20 kitchen 32.64 0.941 0.105 1.80 bonsai 33.26 0.955 0.160 1.30 Table 7: Per-scene quantitative results on the MipNeRF 360 (Barron et al. 2022) dataset u...

work page 2022

[6] [1996]

In Siggraph, volume 96, 43–54

The lumigraph. In Siggraph, volume 96, 43–54. Graves, A.; Bellemare, M. G.; Menick, J.; Munos, R.; and Kavukcuoglu, K. 2017. Automated curriculum learning for neural networks. arXiv preprint arXiv:1704.03003. Hedman, P.; Philip, J.; Price, T.; Frahm, J.-M.; Drettakis, G.; and Brostow, G. 2018. Deep blending for free-viewpoint image-based rendering. In SIG...

arXiv 2017

[7] [2016]

Advances in neural information processing systems , 29

Learning to learn by gradient descent by gradient de- scent. Advances in neural information processing systems , 29. Barron, J. T.; Mildenhall, B.; Verbin, D.; Srinivasan, P. P.; and Hedman, P. 2022. Mip-nerf 360: Unbounded anti- aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5470–5479...

arXiv 2022

[8] [2018]

In Proceedings of the International Conference on Machine Learning (ICML)

Efficient Neural Architecture Search via Parameter Sharing. In Proceedings of the International Conference on Machine Learning (ICML). Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal Policy Optimization Algorithms. arXiv preprint arXiv:1707.06347. Sitzmann, V .; Thies, J.; Heide, F.; Nießner, M.; Wetzstein, G.; and Zollh...

arXiv 2017