📐 GALP — Geometry-Aware Layout Prediction

Geometry-Aware Layout Prediction (GALP) is a core submodule of SceneConductor that predicts an initial 3D scene layout from segmented object masks and corresponding GLB assets.

🔗 Project: https://github.com/jhkim0759/SceneConductor

---

✨ Overview

GALP takes the segmented objects produced in Stage 1 and predicts:

Within the SceneConductor pipeline, GALP serves as the final step of Stage 1.

The generated outputs:

layout_prediction.json
layout_prediction.glb

---

📂 Model Checkpoints

The checkpoints/ directory in this repository is provided as a placeholder.

All pretrained GALP checkpoints can be downloaded from:

🤗 Hugging Face

https://huggingface.co/WopperSet/SceneConductor

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="WopperSet/SceneConductor",
    allow_patterns="checkpoints/*",
    local_dir="./checkpoints"
)

Expected structure:

checkpoints/galp/
├── checkpoint.pt
├── pipeline.yaml
├── galp.yaml
└── condition_embedder.ckpt

For detailed checkpoint organization and download instructions, please refer to the main SceneConductor repository.

---

🛠️ Environment Setup

Environment creation is managed through YAML configuration files.

./setup.sh

Or update the provided YAML according to your system configuration.

---

🚀 Quick Start

Run GALP on a sample scene:

python demo.py --scene assets/0000000 \
                --ckpt  checkpoints/checkpoint.pt \
                --output output/demo_scene.glb \
                --gpu 0

📥 Input

assets/0000000/
├── object_masks/
├── object_glbs/
└── metadata.json

📤 Output

output/
├── demo_scene.glb
└── layout_prediction.json

---

🗄️ Dataset

🚧 Coming Soon

Dataset preparation and download instructions will be released in a future update.

Planned support includes:

• 🪑 3D-FUTURE
• 🏢 ScanNet
• 🖼️ COCO

Stay tuned!

---

🎓 Training

Modify the training scripts according to your environment:

• GPU configuration
• Number of workers
• Dataset paths
• Batch size
• Training hyperparameters

Initial Layout Training

sh script/train_init.sh

Post Refinement Training

sh script/train_post.sh

---

📁 Repository Structure

GALP/
├── assets/
├── checkpoints/
├── configs/
├── script/
├── demo.py
└── README.md

---

🙏 Acknowledgements

This project builds upon several outstanding open-source works:

• SceneGen
• PartCrafter
• SAM3D

We thank the authors for open-sourcing their work.

📜 Citation

If you find GALP useful for your research, please consider citing the corresponding SceneConductor paper.

⭐ Star the repository if it helps your work!