---
license: cc-by-4.0
task_categories:
- text-to-image
language:
- en
tags:
- vision
- evaluation
- diagnostic
- AI-Obedience
pretty_name: VIOLIN
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: test
    path: "violin-test.parquet"
rai:
  dataLimitations: "This dataset is focused on specific color/shape combinations and may not cover all real-world visual scenarios."
  dataBiases: "The images are synthetic and may reflect biases inherent in the generative models used."
  personalSensitiveInformation: "None. The dataset contains no PII (Personally Identifiable Information)."
  dataUseCases: "Research on visual reasoning and color-object association in multimodal models."
  dataSocialImpact: "Provides a benchmark for evaluating model alignment with human-centric visual concepts."
  hasSyntheticData: true
---

<div align="left" style="display: flex; align-items: center; gap: 12px; margin-bottom: 5px;">
  <img src="assets/logo.png" style="height: 36px; width: auto; flex-shrink: 0;">
  <span style="font-size: 32px; font-weight: 700; color: #1f2937; line-height: 1;">
    VIOLIN: VIsual Obedience Level-4 EvaluatIoN
  </span>
</div>

<hr style="height: 1px; border: none; background-color: #e5e7eb; margin-top: 5px; margin-bottom: 15px;">


<p align="center">
  <a href="https://github.com/AI-Obedience/Violin">
    <img src="https://img.shields.io/badge/GitHub-Code-blue?logo=github" alt="GitHub">
  </a>
  <a href="https://arxiv.org/abs/2603.00166">
    <img src="https://img.shields.io/badge/arXiv-2603.00166-B31B1B.svg" alt="arXiv">
  </a>
  <a href="https://ai-obedience.github.io">
    <img src="https://img.shields.io/badge/Project-Page-green" alt="Project Page">
  </a>
</p>


<p align="center">
  <img src="assets/moti.png" alt="Motivation" style="max-width: 100%;">
</p>

**VIOLIN** (VIsual Obedience Level-4 EvaluatIoN) is a diagnostic benchmark designed to assess the **Level-4 Instructional Obedience** of text-to-image generative models.

While state-of-the-art models can render complex semantic scenes (e.g., "Cyberpunk cityscapes"), they often fail at the most fundamental deterministic tasks: generating a perfectly pure, texture-less color image. VIOLIN provides a rigorous framework to measure this "Paradox of Simplicity."



<p align="center">
  <img src="assets/benchmark_display.png" alt="benchmark" style="max-width: 100%;">
</p>

## 📊 Dataset Structure

| Task | Description | Metrics |
| :--- | :--- | :--- |
| **1. Color purity (single block)** | Full-frame uniform color from ISCC–NBS Level-2. | <em>d</em><sub>rgb_ed</sub>, <em>d</em><sub>lab_00</sub>, <em>d</em><sub>sd</sub>, <em>d</em><sub>ced</sub>, <em>d</em><sub>hf</sub> |
| **2. Color purity (two blocks)** | Fill uniform color in two regions. | <em>d</em><sub>rgb_ed</sub>, <em>d</em><sub>lab_00</sub>, <em>d</em><sub>sd</sub>, <em>d</em><sub>ced</sub>, <em>d</em><sub>hf</sub> |
| **3. Geometric shape** | Render a specified shape (e.g., a circle) at a prompt-defined position and scale. | <em>d</em><sub>iou</sub>, <em>d</em><sub>dist</sub>, <em>d</em><sub>size</sub>, <em>d</em><sub>shape</sub>, <em>d</em><sub>pure</sub> |
| **4. Image masking** | Apply mask on image. Based on the image in TencentARC [BrushNet](https://github.com/TencentARC/BrushNet). | <em>d</em><sub>iou</sub>, <em>d</em><sub>biou</sub>, <em>d</em><sub>leak</sub>, <em>d</em><sub>dist</sub>, <em>d</em><sub>edge</sub> |



## 📁 How to Use
You can load the dataset directly via the Hugging Face `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("Perkzi/VIOLIN", split="test")
print(dataset[0])
```


## 📜 Citation

If you find this dataset or our research helpful, please consider citing our paper:

```bibtex
@article{li2026exploring,
  title={Exploring the AI Obedience: Why is Generating a Pure Color Image Harder than CyberPunk?},
  author={Li, Hongyu and Liu, Kuan and Chen, Yuan and Hu, Juntao and Lu, Huimin and Chen, Guanjie and Liu, Xue and Lu, Guangming and Huang, Hong},
  journal={arXiv preprint arXiv:2603.00166},
  year={2026}
}
```