Awesome-3D-Low Level Vision
June 1, 2026 · View on GitHub
R3eVision: A Survey on Robust Rendering, Restoration, and Enhancement for 3D Low-Level Vision
This repository provides a curated collection of papers, benchmarks, and resources from our survey:
"R3eVision: A Survey on Robust Rendering, Restoration, and Enhancement for 3D Low-Level Vision" (arXiv2025).
📝 Authors: Weeyoung Kwon1, Jeahun Sung2, Minkyu Jeon3, Chanho Eom1, and Jihyong Oh2†
🎓 Institution:
- 1 Department of Metaverse Convergence, GSAIM, Chung-Ang University
- 2 Department of Imaging Science, GSAIM, Chung-Ang University
- 3 Department of Computer Science, Princeton University
📘 Abstract
Neural rendering methods such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have achieved significant progress in photorealistic 3D scene reconstruction and novel view synthesis. However, most existing models assume clean and high-resolution (HR) multi-view inputs, which limits their robustness under real-world degradations such as noise, blur, low-resolution (LR), and weather-induced artifacts. To address these limitations, the emerging field of 3D Low-Level Vision (3D LLV) extends classical 2D Low-Level Vision tasks including super-resolution (SR), deblurring, weather degradation removal, restoration, and enhancement into the 3D spatial domain. This survey, referred to as R3eVision, provides a comprehensive overview of robust rendering, restoration, and enhancement for 3D LLV by formalizing the degradation-aware rendering problem and identifying key challenges related to spatio-temporal consistency and ill-posed optimization. Recent methods that integrate LLV into neural rendering frameworks are categorized to illustrate how they enable high-fidelity 3D reconstruction under adverse conditions. Application domains such as autonomous driving, AR/VR, and robotics are also discussed, where reliable 3D perception from degraded inputs is critical. By reviewing representative methods, datasets, and evaluation protocols, this work positions 3D LLV as a fundamental direction for robust 3D content generation and scene-level reconstruction in real-world environments. We maintain an up-to-date project page: https://github.com/CMLab-Korea/Awesome-3D-Low-Level-Vision.
📚 Contents
📣 News
- 📌 2026-04: CVPR 2026 papers updated.
- 📌 2026-02: ICLR 2026 papers updated.
- 📌 2026-01: AAAI 2026 papers updated.
- 📌 2025-06: Paper released to ArXiv.
- 🚀 2025-05: Repository initialized.
🔖 Citation
If you find this survey helpful, please consider citing us:
@article{Kwon2025R3eVision,
title={R3eVision: A Survey on Robust Rendering, Restoration, and Enhancement for 3D Low-Level Vision},
author={Kwon, Weeyoung and Sung, Jeahun and Jeon, Minkyu and Eom, Chanho and Oh, Jihyong},
journal={arXiv preprint arXiv:2506.01061},
year={2025}
}
🧩 Community Contribution
We welcome contributions from the 3D LLV research community!
If you have a new method, dataset, or related resource that fits within the scope of this 3D LLV repository, please feel free to submit a pull request (PR) with the following:
A brief description of your method/resource
Relevant links (e.g., arXiv, project page, code)
Suggested placement (e.g., under “4.2. Deblurring in 3D LLV”)
Our maintainers will review your submission and merge it if appropriate. We hope this page will grow into a collaborative hub for 3D Low-Level Vision (3D LLV) research, covering topics such as degradation-aware rendering, neural field restoration, and robust 3D reconstruction under real-world conditions.
🔍 Survey Paper
You can find the preprint of our survey here:
📄 arXiv:2506.16262
📄 Paper List
We categorize recent 3D LLV papers by methodology (up to February 5, 2026):
4.Low-Level Vision For Robust 3D Rendering
4.1. Super-Resolution (SR) in 3D LLV
4.1.1. Predictive Method
4.1.2. Generative-based
| Title | Publication | Date | Tags |
|---|---|---|---|
| Super-NeRF: View-consistent Detail Generation for NeRF Super-resolution | TVCG | 2024 | Image-based |
| DiSR-NeRF: Diffusion-Guided View-Consistent Super-Resolution NeRF | CVPR | 2024 | Image-based |
| GaussianSR: 3D Gaussian Super-Resolution with 2D Diffusion Priors | arXiv | 2024 | Image-based |
| S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splatting | CVPR | 2025 | Image-based |
| Bridging Diffusion Models and 3D Representations: A 3D Consistent Super-Resolution Framework | ICCV | 2025 | Image-based |
| Arbitrary-Scale 3D Gaussian Super-Resolution | AAAI | 2026 | Image-based |
| SplatSuRe: Selective Super-Resolution for Multi-view Consistent 3D Gaussian Splatting | CVPR | 2026 | Image-based |
| SuperGaussian: Repurposing Video Models for 3D Super Resolution | ECCV | 2024 | Video-based |
4.2. Deblurring in 3D LLV
4.2.1. Motion Deblurring
4.2.2. Motion and Defocus Deblurring
4.3. Weather Degradation Removal on 3D LLV
4.3.1. Pysics-based
| Title | Publication | Date | Tags |
|---|---|---|---|
| DehazeNeRF: Multi-image Haze Removal and 3D Shape Reconstruction using Neural Radiance Fields | 3DV | 2024 | - |
| Dehazing-NeRF: Neural Radiance Fields from Hazy Images | arXiv | 2023 | - |
| ScatterNeRF: Seeing Through Fog with Physically-Based Inverse Neural Rendering | ICCV | 2023 | - |
| DehazeGS: Seeing Through Fog with 3D Gaussian Splatting | arXiv | 2025 | - |
| I2-NeRF: Learning Neural Radiance Fields Under Physically-Grounded Media Interactions | NeurIPS | 2025 | - |
| NimbusGS: Unified 3D Scene Reconstruction under Hybrid Weather | CVPR | 2026 | - |
4.3.2. Detection-based
| Title | Publication | Date | Tags |
|---|---|---|---|
| DerainNeRF: 3D Scene Estimation with Adhesive Waterdrop Removal | ICRA | 2024 | - |
| WeatherGS: 3D Scene Reconstruction in Adverse Weather Conditions via Gaussian Splatting | arXiv | 2024 | - |
| DeRainGS: Gaussian Splatting for Enhanced Scene Reconstruction in Rainy Environments | AAAI | 2025 | - |
| NimbusGS: Unified 3D Scene Reconstruction under Hybrid Weather | CVPR | 2026 | - |
4.4. Restoration
4.4.1. Reference-based
4.4.2. Generative-based
| Title | Publication | Date | Tags |
|---|---|---|---|
| RaFE: Generative Radiance Fields Restoration | ECCV | 2024 | - |
| Drantal-NeRF: Diffusion-Based Restoration for Anti-aliasing Neural Radiance Field | arXiv | 2024 | - |
| ReSplat: Degradation-agnostic Feed-forward Gaussian Splatting via Self-guided Residual Diffusion | ICLR | 2026 | - |
| Geometry-Aware Representation Denoising for Robust Multi-view 3D Reconstruction | arXiv | 2026 | - |
4.5. Enhancement in 3D LLV
4.5.1. Low-light Enhancement
4.5.2. Detail Enhancement
| Title | Publication | Date | Tags |
|---|---|---|---|
| Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields | ICCV | 2021 | Antialiasing-based |
| Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields | CVPR | 2022 | Antialiasing-based |
| Zip-NeRF: Anti-Aliased Grid-Based Neural Radiance Fields | ICCV | 2023 | Antialiasing-based |
| Tri-MipRF: Tri-Mip Representation for Efficient Anti-Aliasing Neural Radiance Fields | ICCV | 2023 | Antialiasing-based |
| Mip-Splatting: Alias-free 3D Gaussian Splatting | CVPR | 2024 | Antialiasing-based |
| BungeeNeRF: Progressive Neural Radiance Field for Extreme Multi-scale Scene Rendering | ECCV | 2022 | Multi-scale-based |
| PyNeRF: Pyramidal Neural Radiance Fields | NeurIPS | 2023 | Multi-scale-based |
| Multiscale Representation for Real-Time Anti-Aliasing Neural Rendering | ICCV | 2023 | Multi-scale-based |
| Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering | CVPR | 2024 | Multi-scale-based |
| HoGS: Unified Near and Far Object Reconstruction via Homogeneous Gaussian Splatting | CVPR | 2025 | Multi-scale-based |
4.5.3. Texture Enhancement
| Title | Publication | Date | Tags |
|---|---|---|---|
| LATTE3D: Large-scale Amortized Text-To-Enhanced3D Synthesis | ECCV | 2024 | Diffusion-based |
| 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors | NeurIPS | 2024 | Diffusion-based |
| 3DEnhancer: Consistent Multi-View Diffusion for 3D Enhancement | CVPR | 2025 | Diffusion-based |
| SEGS-SLAM: Structure-enhanced 3D Gaussian Splatting SLAM with Appearance Embedding | ICCV | 2025 | - |
| Semantic and Rendering Enhancements in 3D Gaussian Modeling with Anisotropic Local Encoding | ICCV | 2025 | - |
5. Experimental Setup
📊 5.1. Datasets & Benchmarks
We include commonly used datasets for evaluating 3D-LLV performance.
| Dataset | Year | Type | #Scenes | #Images | Resolution | Motion |
|---|---|---|---|---|---|---|
| DTU | 2014 | `R$ | 124 | 49 \text{or} 64 | 1600 \times 1200 | $S` |
| Tanks and Temples | 2017 | `R$ | 21 | 100–400 | 1920 \times 1080 | $S` |
| Deep Blending | 2018 | `R$ | 19 | 12–418 | 1288 \times 816 | $S` |
| LLFF | 2019 | `R$ | 8 | 20–62 | 6032 \times 3024 | $S` |
| Stereo Blur Dataset | 2019 | `R$ | 135 | \text{Video} | 1280 \times 720 | $D` |
| NeRF-synthetic | 2020 | `S$ | 8 | 400 | 800 \times 800 | $S` |
| BlendedMVS | 2020 | R/`S$ | 113 | 150–200 | 2048 \times 1536 | $S` |
| NSVF Synthetic | 2020 | `S$ | 8 | 400 | 800 \times 800 | $S` |
| HyperNeRF | 2021 | R/`S$ | 7 | \text{Video} | 1920 \times 1080 | $D` |
| Deblur-NeRF | 2022 | R/`S$ | 31 | 27–53 | 600 \times 400 | $D` |
| NeRF in the Dark | 2022 | R/`S$ | 5 | 25–200 | 6000 \times 4000 | $S` |
| Mip-NeRF 360 | 2022 | `R$ | 9 | 100–330 | 4096 \times 3286 | $S` |
| RTMV | 2022 | `R$ | 2000 | 150 | 1600 \times 1600 | $S` |
| iPhone Dataset | 2022 | `R$ | 14 | \text{Video} | 720 \times 960 | $D` |
| Objaverse | 2023 | S | 800K+ | 3D Object | - | S |
Type: Scene Type (R: Real,S:Synthetic)Motion: Motion Type (S: Static,D: Dynamic)
📈 5.2. Evaluation Metrics
This section summarizes commonly used metrics for evaluating the quality of 3D low-level vision results.
| Title | Publication (Venue / Journal) | Tags | Year |
|---|---|---|---|
| PSNR | - | Full-Rerence | - |
| SSIM | IEEE TIP | Full-Rerence | 2004 |
| LPIPS | CVPR | Full-Rerence | 2018 |
| BRISQUE | IEEE TIP | No-Reference | 2012 |
| NIQE | IEEE SPL | No-Reference | 2013 |
| PIQE | NCC | No-Reference | 2015 |
| RankIQA | ICCV | No-Reference | 2017 |
| MetaIQA | CVPR | No-Reference | 2020 |
| MUSIQ | ICCV | No-Reference | 2021 |
| MANIQA | CVPR | No-Reference | 2022 |
| LIQE | CVPR | No-Reference | 2023 |
| CLIP-IQA | AAAi | No-Reference | 2023 |
| tOF | CVPR | Temporal-Consistency | 2020 |
🆚 Full-reference Metric
These metrics compare each interpolated frame to its ground truth (GT) reference on a pixel level.
-
PSNR (Peak Signal-to-Noise Ratio)
Measures reconstruction fidelity via Mean Squared Error (MSE).
📌 Higher is better, but it often doesn't align with human perception, especially in high-frequency regions. -
SSIM (Structural Similarity Index)
Compares luminance, contrast, and texture to evaluate structural similarity.
📌 More perceptually aligned than PSNR. Higher SSIM indicates stronger similarity. -
LPIPS (Learned Perceptual Image Patch Similarity)
Measures perceptual similarity between image patches using deep network features (e.g., from VGG or AlexNet).
📌 Strong alignment with human perception but dependent on network backbone and training data.
📊 No-reference Metric
These metrics have been proposed to evaluate visual quality without requiring GT data.
-
BRISQUE (Blind/Referenceless Image Spatial Quality Evaluator)
Measures image quality by modeling natural scene statistics in the spatial domain.
📌 No-reference and fast, but less effective on unseen distortions. -
NIQE (Naturalness Image Quality Evaluator)
A no-reference metric using statistical deviations from natural images. (Statistics approach)
📌 Lower NIQE implies higher natural image quality. -
PIQE
Predicts local perceptual quality with high resolution using patch-level IQA and perceptual index training.
📌 Provides fine-grained spatial quality maps; fast, heuristic-based, and perceptually motivated. -
RankIQA
Trains deep networks to rank image quality by learning from relative comparisons, without requiring precise scores.
📌 Efficiently uses pairwise annotations; avoids subjective scoring bias. -
MetaIQA
A meta-learning-based framework for NR-IQA, enabling fast adaptation to new distortion types.
📌 Generalizes better to unseen data but sensitive to meta-training setup. -
MUSIQ (Multi-Scale Image Quality Transformer)
Uses a transformer architecture with multi-scale image patches to assess quality without reference.
📌 Strong performance across diverse datasets; transformer-based scalability. -
MANIQA
Applies a multi-dimension attention mechanism for NR-IQA using transformers and feature fusion.
📌 Excels in capturing complex distortions and perceptual attributes. -
LIQE
Leverages alignment between image content and language descriptions for blind IQA via multitask learning.
📌 Novel paradigm with strong generalization to real-world image distributions. -
CLIP-IQA
Utilizes CLIP's pretrained vision-language embedding to assess perceptual quality via text-image similarity.
📌 Integrates large-scale vision-language models for improved perceptual reasoning.
⏱️ Temporal Consistency
- tOF
Measures how consistent optical flow is across frames.
📌 Lower tOF = smoother motion continuity.