Source-linked AI summary
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
TL;DR
Smoke-degraded images hinder visibility and the cross-view consistency needed for novel view synthesis. GenSmoke-GS combines restoration, dehazing, structurally constrained MLLM enhancement, 3DGS-MCMC optimization, and repeated-run averaging. On the challenge benchmark, it improves quantitative metrics and visual quality over the compared baselines.
Problem
Smoke reduces visibility and weakens the cross-view consistency required by scene optimization and rendering.
Method
GenSmoke-GS uses restoration, dehazing, structurally constrained MLLM enhancement, 3DGS-MCMC optimization, and averaging over repeated runs.
Results
PSNR improves from 11.54 to 20.21, SSIM from 0.597 to 0.729, and LPIPS decreases from 0.705 to 0.446 versus 3DGS.
Takeaways & Limitations
The method improves view synthesis quality under smoke degradation, while structural constraints support stable optimization and averaging reduces unstable local artifacts.
Abstract
from arXiv · showhide
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.
1. Introduction
Smoke degrades both visibility and the cross-view consistency needed for stable novel view synthesis. GenSmoke-GS therefore combines enhancement and optimization stages to improve visibility while limiting structural changes across views.
- Smoke reduces image visibility and the consistency required by optimization and rendering.
- Improved individual-image quality does not necessarily improve novel view synthesis because generative enhancement can introduce structural differences across images.
- GenSmoke-GS uses restoration, dehazing, MLLM-based enhancement with structural constraints, 3DGS-MCMC optimization, and averaging over repeated runs.
- The method aims to improve visibility while keeping enhanced inputs consistent across views.
2. Related Work
Prior work applies radiance-field and Gaussian-splatting methods to degraded scenes, while this method adds preprocessing and enhancement before optimization. The central concern is whether enhanced images remain consistent for rendering.
- Novel view synthesis in degraded scenes: Recent methods address underwater scenes, adverse lighting, and participating media using degradation-related modeling or suitable priors.
- 3D Gaussian Splatting: 3DGS is used for efficient novel view synthesis, with 3DGS-MCMC and FasterGS improving optimization stability and efficiency.
- Relation to our method: GenSmoke-GS differs by enhancing images before optimization in addition to addressing image degradation.
- Relation to our method: The method's overview places restoration, dehazing, MLLM enhancement, 3DGS-MCMC, and repeated-run averaging in sequence.
3. Methods
The method prepares smoke-degraded inputs before scene optimization, constrains generative enhancement to preserve structure, and averages repeated optimized renderings to reduce instability.
- Overview: The pipeline has five stages: preliminary restoration, DCP-based dehazing, MLLM-based enhancement, 3DGS-MCMC optimization, and repeated-run averaging.
- Preliminary restoration and dehazing: ConvIR-UDPNet produces preliminary restored images from smoke-degraded inputs.This stage recovers coarse structures and basic color information.
- Preliminary restoration and dehazing: DCP-based dehazing and preliminary restoration reduce heavy degradation before enhancement.
- MLLM-based enhancement: The MLLM enhances each image independently using a prompt that preserves geometry, layout, object boundaries, and local structures while allowing visibility improvement.
- Optimization and averaging: The enhanced images are optimized with 3DGS-MCMC accelerated by FasterGS, after which novel views are rendered.
- Optimization and averaging: Independent runs are averaged because identical settings can produce slightly different local results; the final submission uses n = 91 runs to reduce unstable local artifacts.
4. Experiment
The method is evaluated on smoke-degraded RealX3D scenes using quantitative metrics and qualitative comparisons against four rendering baselines. It achieves the best average performance across all three metrics, with improvements reported consistently across scenes, while severe degradation remains challenging.
- Benchmark and metrics: The evaluation uses RealX3D smoke-degraded scenes and reports PSNR, SSIM, and LPIPS for novel-view synthesis.The comparisons use provided rendered outputs and corresponding test-set ground-truth views.
- Quantitative results: 20.21 PSNR, 0.729 SSIM, and 0.446 LPIPS are achieved by GenSmoke-GS, improving over 3DGS values of 11.54, 0.597, and 0.705, respectively.Table 2 reports the average results over all evaluated scenes and identifies GenSmoke-GS as best on all three metrics.
- Per-scene and qualitative findings: Improvements are consistent across evaluated scenes, although Shirohana remains relatively difficult under severe degradation.The results indicate improved view synthesis quality under smoke degradation, while repeated runs help reduce unstable local artifacts.
- Qualitative comparisons: The compared methods are 3DGS, I2-NeRF, SeaSplat, SeaThru-NeRF, GenSmoke-GS, and ground truth.Figures 2 and 3 present representative comparisons for Futaba view 0024 and Shirohana view 0027.
5. Conclusion
GenSmoke-GS is a multi-stage method for novel view synthesis from smoke-degraded images in the NTIRE 2026 3DRR Challenge. The results support its effectiveness in the challenge setting, with structural-change control and repeated-run averaging contributing to stable and improved rendered views.
- Method: GenSmoke-GS combines image restoration, dehazing, MLLM-based enhancement, 3DGS-MCMC optimization, and averaging over repeated runs.The method targets novel view synthesis from smoke-degraded images in the NTIRE 2026 3DRR Challenge.
- Results: The experimental results show that GenSmoke-GS is effective for the challenge setting.
- Conclusion: Limiting structural changes during enhancement supports stable optimization, while averaging repeated runs further improves final rendered views.