GFPack++: Attention-Driven Gradient Fields for Optimizing 2D Irregular Packing

Published at ICCV 2025, International Conference on Computer Vision.

Authors

Tianyang Xue¹, Lin Lu¹*, Yang Liu², Mingdong Wu³, Hao Dong³, Yanbin Zhang¹, Renmin Han¹, Baoquan Chen^3
1Shandong University, ²Microsoft Research Asia, ³Peking University

Abstract

2D irregular packing is a classic combinatorial optimization problem with various applications, such as material utilization and texture atlas generation. Due to its NP-hard nature, conventional numerical approaches typically encounter slow convergence and high computational costs. Previous research (GFPack) introduced a generative method for gradient-based packing, providing early evidence of its feasibility but faced limitations such as insufficient rotation support, poor boundary adaptability, and high overlap ratios. In this paper, we propose GFPack++, a deeply investigated framework that adopts attention-based geometry and relation encoding, enabling more comprehensive modeling of complex packing relationships. We further design a constrained gradient and a weighting function to enhance both the feasibility of the produced solutions and the learning effectiveness. Experimental results on multiple datasets demonstrate that GFPack++ achieves higher space utilization, supports continuous rotation, generalizes well to arbitrary boundaries, and infers orders of magnitude faster than previous approaches.

This repository accompanies the ICCV 2025 paper GFPack++: Attention-Driven Gradient Fields for Optimizing 2D Irregular Packing. The code base provides training, evaluation, and sampling utilities for our attention-enhanced score model that learns packing gradients directly from teacher demonstrations. The repository is organised so the Python sources live under src/, shell entry-points under scripts/, and binary extensions / datasets under dedicated folders.
The current release merges the collision potential field and gap-removal enhancement so both operate together during sampling. In our experiments this behaves comparably to the original implementation where the collision field is applied separately from the inference update, while being more efficient because the fused formulation reuses the same intermediate computations.

Directory Layout

.
├── checkpoints/           # Pretrained weights and archived snapshots
├── data/
│   ├── datasets/          # Training / validation .pkl files
│   ├── polygons/          # Polygon definitions (0.txt …)
│   └── gfpack.zip         # Optional compressed resources
├── environment.yml        # Conda environment specification
├── external/
│   └── rmspacing_cpp/     # Native extensions (build scripts + sources)
├── logs/                  # TensorBoard logs written during runs
├── scripts/               # Helper shell / utility scripts
└── src/                   # Python package (`python -m src.train`, etc.)

Key Python modules:

File	Purpose
src/score_model.py	Transformer-based score model
src/gnn_feature.py	Polygon GNN feature extractor
src/geometry_utils.py	Polygon grid utilities + RNG seeding helper
src/sde.py	SDE definitions and samplers
src/train.py	Distributed training entry-point
src/evaluate.py	Evaluation / inference script

1. Setup

conda env create -f environment.yml
conda activate gfpack

All scripts and Python entry-points assume the gfpack environment is active.

2. Build the Native Extensions

The model relies on two pybind11 extensions: calutil and rmspacing. They must be compiled on the target machine before training or evaluation. The rmspacing module enforces workspace bounds (width/height) while resolving overlaps, so keep those parameters aligned with your dataset scale.

cd external/rmspacing_cpp
bash build_util.sh   # builds calutil*.so
bash build.sh        # builds rmspacing*.so

Copy the resulting shared objects into src/ (overwrite the placeholders if you rebuilt them):

cp calutil*.so rmspacing*.so ../../src/

3. Data Preparation

• Download the dataset and ckpt from Google Drive.
• Place training datasets (e.g. dataset_train.pkl) inside data/datasets/. The validation set dataset_vali.pkl is already staged there.
• Polygon definitions (0.txt …) live directly inside data/polygons/. If you have an archive, extract it so every file is available at: data/polygons/<poly_id>.txt.

4. Training

scripts/train.sh wraps torchrun and always executes from the project root. It loads the latest checkpoint from checkpoints/score_model.pth when present.

conda activate gfpack
./scripts/train.sh --log_dir run_001 --batch_size 512

Outputs:

• TensorBoard logs → logs/run_001/train/**
• Latest weights → checkpoints/score_model.pth (and a pickle copy)
• Snapshot per vali_freq epochs → checkpoints/snapshots/score_model_epoch*.pth

All standard argparse flags from src/train.py can be forwarded through the shell script.

5. Evaluation / Sampling

scripts/eval.sh launches the evaluation pipeline (single GPU by default) and expects the trained checkpoint in checkpoints/score_model.pth.

conda activate gfpack
./scripts/eval.sh --log_dir eval_run

Results, including rendered polygon placements, will be written to the same logging hierarchy (logs/eval_run/eval/**). Open TensorBoard to inspect the metrics and figures:

tensorboard --logdir logs

6. Questions / Contributions

Issues and pull requests are welcome. If you hit a build or runtime problem, provide:

1. The exact command you executed (scripts/train.sh …)
2. Conda / CUDA versions
3. A brief excerpt of the console or TensorBoard output

That information will make it much easier to help.

Citation

If you find this project useful, please cite our ICCV 2025 paper:

@inproceedings{Xue2025GFPackPP,
  title     = {GFPack++: Attention-Driven Gradient Fields for Optimizing 2D Irregular Packing},
  author    = {Tianyang Xue and Lin Lu and Yang Liu and Mingdong Wu and Hao Dong and Yanbin Zhang and Renmin Han and Baoquan Chen},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}
@inproceedings{Xue2023GFPack,
title = {Learning Gradient Fields for Scalable and Generalizable Irregular Packing},
author = {Xue, Tianyang and Wu, Mingdong and Lu, Lin and Wang, Haoxuan and Dong, Hao and Chen, Baoquan},
booktitle = {SIGGRAPH Asia 2023 Conference Papers},
year = {2023},
}