pith. machine review for the scientific record. sign in

arxiv: 2604.03039 · v2 · submitted 2026-04-03 · 💻 cs.CV

Recognition: no theorem link

GenSmoke-GS: A Multi-Stage Method for Novel View Synthesis from Smoke-Degraded Images Using a Generative Model

Qida Cao , Xinyuan Hu , Changyue Shi , Jiajun Ding , Zhou Yu , Jun Yu
Authors on Pith no claims yet

Pith reviewed 2026-05-13 20:22 UTC · model grok-4.3

classification 💻 cs.CV
keywords novel view synthesissmoke degradation3D Gaussian splattingimage restorationdehazinggenerative enhancementNTIRE challenge
0
0 comments X

The pith

A multi-stage pipeline of restoration, dehazing, and 3D Gaussian splatting produces clearer novel views from smoke-degraded images.

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

The paper describes a method for generating novel views when input images are obscured by smoke, which reduces visibility and weakens the consistency across viewpoints needed for 3D optimization. It applies a sequence of image restoration, dehazing, multimodal large language model enhancement, 3D Gaussian splatting optimized with MCMC, and averaging across repeated runs. The pipeline aims to increase visibility in each view while keeping the underlying scene content consistent enough for the optimization to succeed. Experiments on the challenge benchmark show higher quantitative scores and improved visual quality compared to baselines. The approach ranked first among 14 entries in the relevant track.

Core claim

The central claim is that a multi-stage pipeline of restoration and enhancement steps followed by 3DGS-MCMC optimization and run averaging can recover sufficient visibility for high-quality novel view synthesis from smoke-degraded inputs while preserving the cross-view consistency required by the 3D optimization process.

What carries the argument

A multi-stage pipeline that performs image restoration and dehazing before applying 3D Gaussian splatting with MCMC optimization, followed by averaging repeated runs to maintain consistency.

Load-bearing premise

The restoration and enhancement steps improve visibility without introducing inconsistent changes to scene content across the different input views.

What would settle it

A direct comparison showing that the final rendered novel views become inconsistent or lower quality when the restoration steps are replaced by ones that alter content differently across views.

Figures

Figures reproduced from arXiv: 2604.03039 by Changyue Shi, Jiajun Ding, Jun Yu, Qida Cao, Xinyuan Hu, Zhou Yu.

Figure 1
Figure 1. Figure 1: Overview of the method. Smoke-degraded images are [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative qualitative comparison on Futaba, view 0024. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Representative qualitative comparison on Shirohana, view 0027. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
read the original abstract

This paper describes our method for Track 2 of the NTIRE 2026 3D Restoration and Reconstruction (3DRR) Challenge on smoke-degraded images. In this task, smoke reduces image visibility and weakens the cross-view consistency required by scene optimization and rendering. We address this problem with a multi-stage pipeline consisting of image restoration, dehazing, MLLM-based enhancement, 3DGS-MCMC optimization, and averaging over repeated runs. The main purpose of the pipeline is to improve visibility before rendering while limiting scene-content changes across input views. Experimental results on the challenge benchmark show improved quantitative performance and better visual quality than the provided baselines. The code is available at https://github.com/plbbl/GenSmoke-GS. Our method achieved a ranking of 1 out of 14 participants in Track 2 of the NTIRE 3DRR Challenge, as reported on the official competition website: https://www.codabench.org/competitions/13993/#/results-tab.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 2 minor

Summary. The paper presents GenSmoke-GS, a multi-stage pipeline for novel view synthesis from smoke-degraded images in Track 2 of the NTIRE 2026 3DRR Challenge. The pipeline combines image restoration, dehazing, MLLM-based enhancement, 3DGS-MCMC optimization, and averaging over repeated runs. Its central claim is that this approach improves visibility while preserving cross-view consistency sufficiently for reliable 3D reconstruction, yielding first-place ranking (1/14) with superior quantitative performance and visual quality over baselines; code is released publicly.

Significance. If the benchmark results hold, the work demonstrates a practical, effective combination of generative restoration and 3D Gaussian Splatting for handling real-world smoke degradation. The top ranking on an official challenge benchmark together with public code release provides a strong, reproducible baseline that can accelerate progress in degraded-image novel-view synthesis.

minor comments (2)
  1. [Abstract] Abstract and §3: the MLLM-based enhancement step is described only at high level; specifying the exact model (e.g., GPT-4V or LLaVA) and the prompt template used would improve reproducibility.
  2. [Experimental Results] §4.2 and Table 1: while the 1/14 ranking is stated, the manuscript does not tabulate the precise PSNR/SSIM/LPIPS values obtained by the method versus the official baselines; adding these numbers would strengthen the quantitative claim.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their thorough review, positive assessment of the work, and recommendation to accept the manuscript. We are pleased that the referee recognizes the practical effectiveness of combining generative restoration techniques with 3D Gaussian Splatting for the smoke-degraded novel view synthesis task, as well as the value of the public code release and the first-place ranking on the official NTIRE 2026 benchmark.

Circularity Check

0 steps flagged

No significant circularity; empirical benchmark ranking stands independently

full rationale

The paper describes a multi-stage pipeline (restoration, dehazing, MLLM enhancement, 3DGS-MCMC, averaging) evaluated solely via external NTIRE 3DRR Challenge Track 2 ranking (1/14). No equations, fitted parameters, or self-citations are presented that reduce the reported performance or cross-view consistency claim to the target result by construction. The central claim remains an independent empirical outcome on a public benchmark with released code, satisfying the criteria for a self-contained non-circular finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No mathematical axioms, free parameters, or invented physical entities are described; the work is an empirical engineering pipeline built from standard vision components.

pith-pipeline@v0.9.0 · 5497 in / 1100 out tokens · 66161 ms · 2026-05-13T20:22:57.538847+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 8 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Dehaze-then-Splat: Generative Dehazing with Physics-Informed 3D Gaussian Splatting for Smoke-Free Novel View Synthesis

    cs.CV 2026-04 unverdicted novelty 5.0

    Dehaze-then-Splat uses per-frame generative dehazing followed by physics-regularized 3D Gaussian Splatting to achieve 20.98 dB PSNR and 0.683 SSIM on the Akikaze scene, a 1.5 dB gain over baseline by mitigating cross-...

  2. 3D Smoke Scene Reconstruction Guided by Vision Priors from Multimodal Large Language Models

    cs.CV 2026-04 unverdicted novelty 5.0

    A framework that combines MLLM-based image enhancement with a medium-aware 3D Gaussian Splatting model to reconstruct and render smoke scenes.

  3. CLIP-Guided Data Augmentation for Night-Time Image Dehazing

    cs.CV 2026-04 unverdicted novelty 5.0

    CLIP-guided selection of external data plus staged NAFNet training and inference fusion provides an effective pipeline for nighttime image dehazing in the NTIRE 2026 challenge.

  4. Training-Free Model Ensemble for Single-Image Super-Resolution via Strong-Branch Compensation

    cs.CV 2026-04 unverdicted novelty 4.0

    A dual-branch training-free ensemble fuses a hybrid attention network with a Mamba-based model via weighted combination to enhance super-resolution PSNR on DIV2K x4.

  5. Dual-Branch Remote Sensing Infrared Image Super-Resolution

    cs.CV 2026-04 unverdicted novelty 4.0

    Dual-branch fusion of HAT-L and MambaIRv2-L with eight-way ensemble and equal-weight averaging outperforms single branches on PSNR, SSIM, and challenge score for infrared super-resolution.

  6. SmokeGS-R: Physics-Guided Pseudo-Clean 3DGS for Real-World Multi-View Smoke Restoration

    cs.CV 2026-04 conditional novelty 4.0

    SmokeGS-R uses refined dark channel prior for pseudo-clean supervision to train 3DGS geometry, followed by ensemble-based appearance harmonization, achieving PSNR 15.21 and outperforming baselines on smoke restoration...

  7. Beyond Model Design: Data-Centric Training and Self-Ensemble for Gaussian Color Image Denoising

    cs.CV 2026-04 unverdicted novelty 3.0

    Expanding training data diversity, adopting two-stage optimization, and applying geometric self-ensemble raises Restormer performance on Gaussian color denoising at sigma=50 by 3.366 dB PSNR on the NTIRE 2026 validation set.

  8. NTIRE 2026 3D Restoration and Reconstruction in Real-world Adverse Conditions: RealX3D Challenge Results

    cs.CV 2026-04 unverdicted novelty 2.0

    The NTIRE 2026 challenge reports measurable progress in 3D reconstruction pipelines that handle real-world low-light and smoke degradation via the RealX3D benchmark.

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · cited by 8 Pith papers

  1. [1]

    Gc-gs: Gradient control gaussian splatting with various image degradation

    Qida Cao, Jiajun Ding, Qingyuan Tang, Tianning Zhao, Xi- aoling Gu, Jianping Fan, and Zhou Yu. Gc-gs: Gradient control gaussian splatting with various image degradation. Pattern Recognition, page 112304, 2025. 1

    work page 2025

  2. [2]

    Revitalizing convolutional network for image restoration

    Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Revitalizing convolutional network for image restoration. IEEE Transactions on Pattern Analysis and Machine Intel- ligence, 46(12):9423–9438, 2024. 2

    work page 2024

  3. [3]

    Luminance-gs: Adapting 3d gaussian splatting to chal- lenging lighting conditions with view-adaptive curve adjustment

    Ziteng Cui, Xuangeng Chu, and Tatsuya Harada. Luminance-gs: Adapting 3d gaussian splatting to chal- lenging lighting conditions with view-adaptive curve adjustment. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 26472–26482, 2025. 1

    work page 2025

  4. [4]

    Faster-gs: Analyzing and improv- ing gaussian splatting optimization.arXiv preprint arXiv:2602.09999, 2026

    Florian Hahlbohm, Linus Franke, Martin Eisemann, and Marcus Magnor. Faster-gs: Analyzing and improv- ing gaussian splatting optimization.arXiv preprint arXiv:2602.09999, 2026. 1, 2

    work page arXiv 2026

  5. [5]

    Single image haze removal using dark channel prior.IEEE transactions on pat- tern analysis and machine intelligence, 33(12):2341–2353,

    Kaiming He, Jian Sun, and Xiaoou Tang. Single image haze removal using dark channel prior.IEEE transactions on pat- tern analysis and machine intelligence, 33(12):2341–2353,

    work page

  6. [6]

    Sr- splat: Feed-forward super-resolution gaussian splatting from sparse multi-view images

    Xinyuan Hu, Changyue Shi, Chuxiao Yang, Minghao Chen, Jiajun Ding, Tao Wei, Chen Wei, Zhou Yu, and Min Tan. Sr- splat: Feed-forward super-resolution gaussian splatting from sparse multi-view images. InProceedings of the AAAI Con- ference on Artificial Intelligence, pages 4950–4958, 2026. 1

    work page 2026

  7. [7]

    3d gaussian splatting for real-time radiance field rendering.ACM Trans

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, George Drettakis, et al. 3d gaussian splatting for real-time radiance field rendering.ACM Trans. Graph., 42(4):139–1,

    work page

  8. [8]

    3d gaussian splat- ting as markov chain monte carlo.Advances in Neural Infor- mation Processing Systems, 37:80965–80986, 2024

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splat- ting as markov chain monte carlo.Advances in Neural Infor- mation Processing Systems, 37:80965–80986, 2024. 1, 2

    work page 2024

  9. [9]

    Seathru- nerf: Neural radiance fields in scattering media

    Deborah Levy, Amit Peleg, Naama Pearl, Dan Rosenbaum, Derya Akkaynak, Simon Korman, and Tali Treibitz. Seathru- nerf: Neural radiance fields in scattering media. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 56–65, 2023. 1, 3

    work page 2023

  10. [10]

    Realx3d: A physically-degraded 3d benchmark for multi-view visual restoration and recon- struction.arXiv preprint arXiv:2512.23437, 2025

    Shuhong Liu, Chenyu Bao, Ziteng Cui, Yun Liu, Xuangeng Chu, Lin Gu, Marcos V Conde, Ryo Umagami, Tomohiro Hashimoto, Zijian Hu, et al. Realx3d: A physically-degraded 3d benchmark for multi-view visual restoration and recon- struction.arXiv preprint arXiv:2512.23437, 2025. 2

    work page arXiv 2025

  11. [11]

    I2-nerf: Learning neural radiance fields un- der physically-grounded media interactions.arXiv preprint arXiv:2510.22161, 2025

    Shuhong Liu, Lin Gu, Ziteng Cui, Xuangeng Chu, and Tat- suya Harada. I2-nerf: Learning neural radiance fields un- der physically-grounded media interactions.arXiv preprint arXiv:2510.22161, 2025. 1, 3

    work page arXiv 2025

  12. [12]

    OpenAI: GPT-Image-1.5.https://openai

    OpenAI. OpenAI: GPT-Image-1.5.https://openai. com / index / new - chatgpt - images - is - here/,

    work page

  13. [13]

    Seasplat: Representing underwater scenes with 3d gaussian splatting and a physically grounded image formation model

    Daniel Yang, John J Leonard, and Yogesh Girdhar. Seasplat: Representing underwater scenes with 3d gaussian splatting and a physically grounded image formation model. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 7632–7638. IEEE, 2025. 1, 3

    work page 2025

  14. [14]

    Lita-gs: Illumination- agnostic novel view synthesis via reference-free 3d gaus- sian splatting and physical priors

    Han Zhou, Wei Dong, and Jun Chen. Lita-gs: Illumination- agnostic novel view synthesis via reference-free 3d gaus- sian splatting and physical priors. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21580–21589, 2025. 1

    work page 2025

  15. [15]

    Udpnet: Unleashing depth-based priors for robust image de- hazing.arXiv preprint arXiv:2601.06909, 2026

    Zengyuan Zuo, Junjun Jiang, Gang Wu, and Xianming Liu. Udpnet: Unleashing depth-based priors for robust image de- hazing.arXiv preprint arXiv:2601.06909, 2026. 2 4

    work page arXiv 2026