---

📌 Overview

GenSmoke-GS is a reconstruction-oriented multi-stage pipeline designed to improve 3D reconstruction under smoke-degraded multi-view conditions.

It consists of two main modules:

• 🔧 2D Enhancement: UDPNet restoration → DCP dehazing → MLLM enhancement
• 🧱 3D Reconstruction: 3DGS-MCMC training → multi-run fusion

---

🖼️ Method Pipeline

---

🔄 Pipeline Summary

Haze Images
   ↓
UDPNet Restoration
   ↓
DCP Dehazing
   ↓
MLLM Enhancement
   ↓
3DGS-MCMC Reconstruction (91 runs)
   ↓
Multi-run Averaging
   ↓
Final NVS Results

---

📂 Repository Structure

2d_enhancement/     # UDPNet + DCP + MLLM enhancement
3d_reconstruction/  # 3DGS-MCMC + FasterGS + fusion
pip.png
README.md

---

🚀 Quick Start

1️⃣ 2D Enhancement (UDPNet + DCP)

cd 2d_enhancement

pip install -r requirements.txt
pip install -r dcp_dehaze/requirements.txt

python run_udp_then_dcp.py \
  --scene_roots <input_dataset_dirs> \
  --out_root <output_dir> \
  --ckpt code/UDPNet/UDPNet_checkpoints/ConvIR_UDPNet_ITS.ckpt

---

📦 External Dependencies & Weights

This module relies on several external methods.

• UDPNet (ConvIR + UDPNet) 🔗 https://github.com/Harbinzzy/UDPNet

• DepthAnything V2 🔗 https://github.com/DepthAnything/Depth-Anything-V2

• Dark Channel Prior (DCP) 🔗 https://github.com/joyeecheung/dark-channel-prior-dehazing

---

📊 Our Released Results (After UDPNet + DCP)

• 🔗 UDPNet outputs: https://pan.baidu.com/s/1Ea5j3WNVK3vdZU8eVMRcAg

• 🔗 DCP dehazed results: https://pan.baidu.com/s/1IDokNAZgEUQw1S2c8iXbFw

---

🧠 MLLM Enhancement (GPT-Image-1.5)

This stage performs view-wise enhancement using the OpenAI image generation API.

---

🔧 Model

• Model: gpt-image-1.5
• Docs: https://platform.openai.com/docs/guides/image-generation

---

📄 Prompt

• Default prompt: gpt_image_prompt_default.txt

---

📊 Our Released Results (MLLM Outputs)

• 🔗 https://pan.baidu.com/s/1M14Tw5RY42ovroslUz0PWA

---

🧱 3️⃣ 3D Reconstruction

cd 3d_reconstruction

bash install.sh
sudo apt install -y colmap

./train.sh /path/to/dataset_parent

---

📊 Our Released Results (3D Reconstruction)

• 🔗 Final NVS results: https://pan.baidu.com/s/1pW--LhgjuKOiCylqCLJcaQ

• 🔗 Multi-run results (91 runs): https://pan.baidu.com/s/1roCxrpJEd8pTqFOMCbMlyQ

---

🧠 Gaussian Weights (Released)

We additionally provide trained Gaussian splatting weights from runs 1–20:

• 🔗 https://pan.baidu.com/s/12P0-5X9deoqdl-TMe1fY2g?pwd=plbb

📌 Notes:

• These weights are obtained from 20 independent runs

• The performance is comparable to the full 91-run averaging results

• Useful for:

  • faster evaluation
  • lightweight reproduction
  • ablation studies

---

📁 Dataset

• 🔗 https://pan.baidu.com/s/1IoIEuE9XTFhEb4meOYp3_g?pwd=plbb

---

🚧 Release Plan

✅ 2026-03-25

• Results released

✅ 2026-03-26

• Code released

✅ 2026-03-27

• Gaussian weights (partial) released

---

📦 TODO

• Upload results
• Upload code
• Upload Gaussian weights (1–20 runs)

---

📬 Contact

If you have any questions:

• Open an Issue
• Contact the authors

---