MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning

📄 Paper

MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning — arXiv:2505.20772

✅ Accepted as a Poster Paper at NeurIPS 2025 🎉

🔥 Introduction

MetaSlot is a novel aggregation module for Object-Centric Learning (OCL) that overcomes two long-standing limitations of conventional Slot Attention models:

🚫 Fixed number of slots
🎲 Random slot initialization

Our approach introduces:

A global vector-quantized (VQ) prototype codebook
A two-stage aggregate-and-deduplicate framework

Together, they enable more adaptive, robust, and interpretable slot representations.

🗂️ Repository Structure


MetaSlot/
│── object_centric_bench/model/metaslot.py   # Core implementation of MetaSlot
│── configs/                                 # Example configs
│    ├── dinosaur_r-voc.py                   # DINOSAUR with MetaSlot
│    ├── vqvae-voc-c4.py                     # VQ-VAE pretraining
│    ├── slotdiffusion_r_vqvae-voc.py        # SlotDiffusion with MetaSlot
│    └── slate_r_vqvae-voc.py                # SLATE with MetaSlot
│── train.py                                 # Training script

Converted Datasets 🚀

Converted datasets, including ClevrTex, COCO, VOC and MOVi-D are available as releases.

dataset-clevrtex: converted dataset ClevrTex.
dataset-coco: converted dataset COCO.
dataset-voc: converted dataset VOC.
dataset-movi_d: converted dataset MOVi-D.

🚀 Getting Started

Install dependencies

conda create -n MetaSlot python=3.10
conda activate MetaSlot
pip install -r requirements.txt

🏋️ Train

1. Training with DINOSAUR + MetaSlot

python train.py \
    --data_dir ./data \
    --cfg_file ./Config/config-metaslot/dinosaur_r-voc.py

2. Training SlotDiffusion + MetaSlot (requires VQ-VAE pretraining)

# Step 1: Train VQ-VAE
python train.py \
    --data_dir ./data \
    --cfg_file ./Config/config-metaslot/vqvae-voc-c4.py

# Step 2: Train SlotDiffusion with pretrained VQ-VAE
python train.py \
    --data_dir ./data \
    --cfg_file ./Config/config-metaslot/slotdiffusion_r_vqvae-voc.py \
    --ckpt_file {your_vqvae_best_ckpt.pth}

3. Training SLATE + MetaSlot (requires VQ-VAE pretraining)

# Step 1: Train VQ-VAE
python train.py \
    --data_dir ./data \
    --cfg_file ./Config/config-metaslot/vqvae-voc-c256.py

# Step 2: Train SLATE with pretrained VQ-VAE
python train.py \
    --data_dir ./data \
    --cfg_file ./Config/config-metaslot/slate_r_vqvae-voc.py \
    --ckpt_file {your_vqvae_best_ckpt.pth}

🔍 Eval

python eval.py \
    --data_dir ./data \
    --cfg_file ./Config/config-metaslot/{your_model_config.py} \
    --ckpt_file {your_ckpt.pth}

🧩 Integration

MetaSlot is plug-and-play 🔌 and can be seamlessly integrated into most OCL pipelines. Simply replace the original aggregator with:

from object_centric_bench.model.metaslot import MetaSlot
aggregator = MetaSlot(...)

📊 Results (Teaser)

The checkpoint for MetaSlot-DINOSAUR on the MS COCO dataset can be downloaded here.

alt text

Acknowledgments

We thank the authors of the following projects for making their code open source:

VQ-VFM-OCL

Citation

If you find our work useful, please cite:

BibTeX

@article{liu2025metaslot,
  title={MetaSlot: Break Through the Fixed Number of Slots in Object-Centric Learning},
  author={Liu, Hongjia and Zhao, Rongzhen and Chen, Haohan and Pajarinen, Joni},
  journal={arXiv preprint arXiv:2505.20772},
  year={2025}
}

📬 Contact

I am a master's student with research interests in representation learning and robotic manipulation.

For questions or potential collaborations, please feel free to open an issue or reach out via email:

📧 hongjia.liu@aalto.fi

MetaSlot

Install / Use

README