YOLO-Count: Differentiable Object Counting for Text-to-Image Generation

Guanning Zeng · Xiang Zhang · Zirui Wang · Haiyang Xu · Zeyuan Chen · Bingnan Li · Zhuowen Tu

ICCV 2025

Paper | arXiv

---

This repository contains the official implementation of YOLO-Count, a fully differentiable and open-vocabulary object counting model. YOLO-Count is designed to provide accurate object count estimation and enable fine-grained quantity control for text-to-image (T2I) generation models.

---

Environment Preparation

We recommend using Conda to set up the environment.

conda create -n yolocnt python=3.12
conda activate yolocnt
pip install -r requirements.txt

---

Dataset Preparation

YOLO-Count is trained and evaluated on multiple object counting benchmarks. Please download and organize each dataset as follows.

FSC147

• Download FSC147 from
  https://github.com/cvlab-stonybrook/LearningToCountEverything
• Place the following folders under:

  data/FSC/
  ├── gt_density_map_adaptive_384_VarV2
  └── images_384_VarV2
  

Open Images v7 (OImgv7)

Download Open Images v7 using:

python -m scripts.download_oimgv7

Objects365 (Obj365)

Download the validation images with:

python -m scripts.download_o365

Then organize the data as:

data/Obj365/objects365/val

LVIS

• Download LVIS
• Place all files under:

  data/LVIS/
  

---

Pre-trained Weights

Pre-trained model weights are available at
https://huggingface.co/zx1239856/yolo-count/tree/main

Please download the weights and place them in the checkpoints/ directory.

---

Evaluation

Evaluation can be performed using the eval_*.py scripts in the scripts folder.
For example, to evaluate on FSC147:

python -m scripts.eval_fsc

Results

The table below reports counting performance using Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE).

Dataset	Split	MAE	RMSE
FSC	Test	15.6745	96.3807
FSC	Validation	14.8297	59.6979
LVIS	Validation	1.5379	5.6076
OImgv7	Validation	3.7087	12.0285
Obj365	Validation	3.2749	9.2181

---

Citation

If you find this work useful in your research, please consider citing:

@InProceedings{zeng2025yolocount,
    author    = {Zeng, Guanning and Zhang, Xiang and Wang, Zirui and Xu, Haiyang and Chen, Zeyuan and Li, Bingnan and Tu, Zhuowen},
    title     = {YOLO-Count: Differentiable Object Counting for Text-to-Image Generation},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {16765--16775}
}

---

License

This repository is released under the CC-BY-SA 4.0 license.