---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: condition
    dtype: image
  - name: caption
    dtype: string
  - name: pose_map
    sequence:
      sequence: float32
  splits:
  - name: train
    num_bytes: 3788921789.965
    num_examples: 13203
  - name: val
    num_bytes: 329702646.598
    num_examples: 1154
  download_size: 4129152495
  dataset_size: 4118624436.563
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: val
    path: data/val-*
task_categories:
- image-to-image
---

# realsinglehumanpose

This dataset is associated with the paper [LISA: Likelihood Score Alignment for Visual-condition Controllable Generation](https://huggingface.co/papers/2606.27192).

- **Repository:** https://github.com/HKUST-LongGroup/LISA
- **Paper:** https://huggingface.co/papers/2606.27192

## Dataset Description

The `realsinglehumanpose` dataset is designed for pose-guided image generation tasks and was evaluated using the LISA framework for controllable image generation.

### Dataset Structure

The dataset contains the following features:
- `image`: The target ground-truth human image.
- `condition`: The input condition image (e.g., pose visual condition).
- `caption`: Text description/caption of the image.
- `pose_map`: Associated pose keypoint coordinates.

## Citation

```bibtex
@misc{wang2026lisalikelihoodscorealignment,
      title={LISA: Likelihood Score Alignment for Visual-condition Controllable Generation}, 
      author={Yanghao Wang and Hongxu Chen and Jiazhen Liu and Zhenqi He and Rui Liu and Zhen Wang and Long Chen},
      year={2026},
      eprint={2606.27192},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.27192}, 
}
```