Hugging Face Community Releases Open Image Preference Dataset
AI Impact Summary
The 🤗 Community has released an open preference dataset for text-to-image generation, focusing on diverse prompt categories and complexities to improve model training. This dataset utilizes synthetic data generation with distilabel and leverages Flux and Stable Diffusion models to create image pairs for preference learning. The dataset’s creation involved filtering for toxicity, manual review, and a model-finetune experiment, highlighting the potential for adapting models through targeted fine-tuning based on preference data.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- info