Perkzi/VIOLIN

VIOLIN: Visual Instruction-based Color Evaluation

---

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."

🧪 Key Scientific Insights

Our research identifies two primary obstacles in current generative AI:

• Aesthetic Inertia: The tendency of models to prioritize visual richness and textures over strict instructional adherence, even when "pure color" or "no texture" is explicitly requested.
• Semantic Gravity: The bias where models follow instructions better when they align with common visual knowledge but fail when context is random or conflicting.

📊 Dataset Structure

The dataset comprises over 42,000 text-image pairs across 6 variations:

Variation	Description	Evaluation Focus
Variation 1	Single Color Block	Basic pixel-level precision (ISCC-NBS)
Variation 2	Two-block Split	Spatial layout and vertical/horizontal split
Variation 3	Four-quadrant Split	Complex spatial reasoning and contrast
Variation 4	Fuzzy Color	Bounded constraints and flexibility
Variation 5	Multilingual	Robustness across English, Chinese, and French
Variation 6	Color Spaces	Cross-format understanding (Hex, RGB, HSL)

📐 Evaluation Metrics

We propose a dual-metric approach for evaluating "Minimum Viable Obedience":

1. Color Precision: Measuring the ΔE (CIEDE2000) or Euclidean distance between the generated pixels and the ground truth.
2. Color Purity: Assessing the presence of artifacts, gradients, or unintended textures using variance-based analysis.

📁 How to Use

You can load the dataset directly via the Hugging Face datasets library:

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:

@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}
}