--- license: apache-2.0 dataset_info: features: - name: image dtype: image - name: internal_id dtype: string - name: prompt dtype: string - name: url dtype: string - name: annotation struct: - name: symmetry dtype: int64 range: [-1,1] - name: richness dtype: int64 range: [-2,2] - name: color aesthetic dtype: int64 range: [-1,1] - name: detail realism dtype: int64 range: [-3,1] - name: safety dtype: int64 range: [-3,1] - name: body dtype: int64 range: [-4,1] - name: lighting aesthetic dtype: int64 range: [-1,2] - name: lighting distinction dtype: int64 range: [-1,2] - name: background dtype: int64 range: [-1,2] - name: emotion dtype: int64 range: [-2,2] - name: main object dtype: int64 range: [-1,1] - name: color brightness dtype: int64 range: [-1,1] - name: face dtype: int64 range: [-3,2] - name: hands dtype: int64 range: [-4,1] - name: clarity dtype: int64 range: [-2,2] - name: detail refinement dtype: int64 range: [-4,2] - name: unsafe type dtype: int64 range: [0,3] - name: object pairing dtype: int64 range: [-1,1] - name: meta_result sequence: dtype: int64 - name: meta_mask sequence: dtype: int64 config_name: default splits: - name: train num_examples: 40743 --- # VisionRewardDB-Image ## Introduction VisionRewardDB-Image is a comprehensive dataset designed to train VisionReward-Image models, providing detailed aesthetic annotations across 18 aspects. The dataset aims to enhance the assessment and understanding of visual aesthetics and quality. 🌟✨ For more detail, please refer to the [**Github Repository**](https://github.com/THUDM/VisionReward). 🔍📚 ## Annotation Detail Each image in the dataset is annotated with the following attributes:
Dimension Attributes
Composition Symmetry; Object pairing; Main object; Richness; Background
Quality Clarity; Color Brightness; Color Aesthetic; Lighting Distinction; Lighting Aesthetic
Fidelity Detail realism; Detail refinement; Body; Face; Hands
Safety & Emotion Emotion; Safety
### Example: Scene Richness (richness) - **2:** Very rich - **1:** Rich - **0:** Normal - **-1:** Monotonous - **-2:** Empty For more detailed annotation guidelines(such as the meanings of different scores and annotation rules), please refer to: - [annotation_deatil](https://flame-spaghetti-eb9.notion.site/VisionReward-Image-Annotation-Detail-196a0162280e80ef8359c38e9e41247e) - [annotation_deatil_zh](https://flame-spaghetti-eb9.notion.site/VisionReward-Image-195a0162280e8044bcb4ec48d000409c) ## Additional Feature Detail The dataset includes two special features: `annotation` and `meta_result`. ### Annotation The `annotation` feature contains scores across 18 different dimensions of image assessment, with each dimension having its own scoring criteria as detailed above. ### Meta Result The `meta_result` feature transforms multi-choice questions into a series of binary judgments. For example, for the `richness` dimension: | Score | Is the image very rich? | Is the image rich? | Is the image not monotonous? | Is the image not empty? | |-------|------------------------|-------------------|---------------------------|----------------------| | 2 | 1 | 1 | 1 | 1 | | 1 | 0 | 1 | 1 | 1 | | 0 | 0 | 0 | 1 | 1 | | -1 | 0 | 0 | 0 | 1 | | -2 | 0 | 0 | 0 | 0 | Each element in the binary array represents a yes/no answer to a specific aspect of the assessment. For detailed questions corresponding to these binary judgments, please refer to the `meta_qa_en.txt` file. ### Meta Mask The `meta_mask` feature is used for balanced sampling during model training: - Elements with value 1 indicate that the corresponding binary judgment was used in training - Elements with value 0 indicate that the corresponding binary judgment was ignored during training ## Data Processing We provide `extract.py` for processing the dataset into JSONL format. The script can optionally extract the balanced positive/negative QA pairs used in VisionReward training by processing `meta_result` and `meta_mask` fields. ```bash python extract.py [--save_imgs] [--process_qa] ``` ## Citation Information ``` @misc{xu2024visionrewardfinegrainedmultidimensionalhuman, title={VisionReward: Fine-Grained Multi-Dimensional Human Preference Learning for Image and Video Generation}, author={Jiazheng Xu and Yu Huang and Jiale Cheng and Yuanming Yang and Jiajun Xu and Yuan Wang and Wenbo Duan and Shen Yang and Qunlin Jin and Shurun Li and Jiayan Teng and Zhuoyi Yang and Wendi Zheng and Xiao Liu and Ming Ding and Xiaohan Zhang and Xiaotao Gu and Shiyu Huang and Minlie Huang and Jie Tang and Yuxiao Dong}, year={2024}, eprint={2412.21059}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2412.21059}, } ```