SCANEDIT Inference

📋 Overview

SCANEDIT is a 3D scene editing framework that uses language models to understand and execute scene manipulation tasks based on text prompts.

✅ Project Status

Completed ✓

• Inference code released
• Installation scripts
• Documentation

TODO 🚧

• Inference sample
• Evaluation setup
• Evaluation data

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/aminebdj/ScanEdit
cd ScanEdit
# Run the installation script
chmod +x install_environment.sh
./install_environment.sh
# Activate the environment
conda activate scanedit

Set up API Key

export GROQ_API_KEY="your_api_key_here"

💻 Usage

Complete Example

SCENE_NAME="3e8bba0176"
python inference.py \
--scene_name "$SCENE_NAME" \
--text_prompt "Create a seating area for a lecture on the whiteboard" \
--path_to_ply "/data/scannetpp/plys/validation/${SCENE_NAME}.ply" \
--masks_classes_path "/data/scannetpp/semantic_3d_masks/${SCENE_NAME}.pt" \
--path_to_rgb_data "/data/scannetpp/data/${SCENE_NAME}/dslr/undistorted_images" \
--path_to_2d_masks "/data/scannetpp/undistorted_rast_masks/${SCENE_NAME}.pt" \
--path_to_annotations "/data/scannetpp/vlm_annotations/${SCENE_NAME}.yaml" \
--dataset "scannetpp" \
--llm_name "openai/gpt-oss-120b" \
--folder_name "results" \
--output_dir "./outputs" \
--verbose

🔧 Arguments Reference

Click to expand full argument list

Paths

• --path_to_rgb_data: RGB images directory
• --path_to_2d_masks: 2D projection masks (.pt)
• --path_to_annotations: VLM annotations (.yaml)
• --path_to_save_additional_masks: Output path for masks
• --img_sample_path: Sample image for testing

Model Configuration

• --groq_api_key: GROQ API key (default: env var)
• --llm_name: LLM model name (default: "openai/gpt-oss-120b")
• --dataset: Dataset type (choices: scannetpp, replica)

Output Settings

• --folder_name: Output subdirectory name
• --output_dir: Base output directory
• --device: Computation device (cuda/cpu)

Flags

• --use_preds: Use predicted masks (adds extension)
• --verbose: Enable detailed logging

📁 Dataset Structure

project_root/
├── data/
│ └── scannetpp/
│ ├── data/
│ │ └── {scene_name}/
│ │ └── dslr/
│ │ └── undistorted_images/
│ │ ├── DSC00001.JPG
│ │ └── ...
│ ├── plys/
│ │ └── validation/
│ │ ├── 3e8bba0176.ply
│ │ └── ...
│ ├── undistorted_rast_masks/
│ │ └── {scene_name}.pt
│ │
│ ├── semantic_3d_masks/
│ │ └── {scene_name}.pt
│ │ │ └── vlm_annotations/
│ └── {scene_name}.yaml
│
└── outputs/
└── {folder_name}/
└── {scene_name}/
├── final.ply
└── ...

📚 Citation

@inproceedings{el2025scanedit,
title={ScanEdit: Hierarchically-Guided Functional 3D Scan Editing},
author={El Amine Boudjoghra, Mohamed and Laptev, Ivan and Dai, Angela},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={27105--27115},
year={2025}
}