Skip to content

Dataset · UCSC-VLAA

197 thousand image editing pairs by instruction

UCSC-VLAA/HQ-Edit

A high-resolution synthetic dataset with detailed text instructions for training image editing models.

What's inside

The dataset includes 197,350 high-resolution image editing examples with detailed English text prompts (metadata indicates a size range from 1K to 10K). The data was collected automatically using a pipeline based on GPT-4V and DALL-E 3 models without manual annotation. The images feature high detail and are accompanied by extensive edit descriptions.

What you can do with it

  • Training and LoRA fine-tuning of diffusion models for text-based image editing
  • Training multimodal models to understand complex image modification instructions
  • Evaluating generative inpainting and local editing quality
Languages
English
Size
1K–10K rows
Downloads this month
4.4K
Likes
42

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("UCSC-VLAA/HQ-Edit")