Dataset · osunlp
Instructions for image editing
osunlp/MagicBrush
10K triplets of source image, instruction, and result for training editing.
What's inside
MagicBrush is the first large-scale instruction-based image editing dataset with manual annotation. It includes 10K triplets (source image, instruction, target image) and covers single-turn and multi-turn editing, with and without a mask. The instruction language is English.
What you can do with it
- Train an instruction-based image editing model
- Fine-tuning LoRA for multi-turn editing
- Test editing with and without a mask
- Tasks
- Text → image, Image → image
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 15.2K
- Likes
- 117
Models trained on it
How to load it
from datasets import load_dataset
ds = load_dataset("osunlp/MagicBrush")