OSU-NLP-Group/MagicBrush
A 10K-sample manually annotated dataset for training and evaluating instruction-guided image editing models using diffusion.

Velocity · 7d
+0.4
★ / day
Trend
→steady
star history
MagicBrush is a large-scale dataset of image triplets (source, instruction, target) supporting instruction-guided image editing. It covers scenarios including single-turn, multi-turn, mask-provided, and mask-free editing. The repository includes training code and a fine-tuned checkpoint for the InstructPix2Pix model. The dataset is published on Huggingface and was accepted to NeurIPS 2023 Datasets and Benchmarks.