MagicBrush
MagicBrush is a manually annotated guided image editing dataset, covering various single-turn and multi-turn editing scenarios, with or without masks. Ideal for training specialized models.
10,000 triples (source image, instruction, target image), PNG/JPEG and text formats
CC-BY 4.0
Description
The dataset MagicBrush includes approximately 10,000 triples consisting of a source image, an editing instruction, and the resulting target image. Manual annotations ensure the quality of edits in various contexts, including edits with or without masks.
What is this dataset for?
- Train image-editing models guided by instructions
- Develop AI assistants for interactive photo editing
- Study single-turn and multi-turn visual editing scenarios
Can it be enriched or improved?
Yes, we can extend this dataset with additional annotations on the nature of the changes or add more multi-turn scenarios. The integration of more precise masks or context metadata would be a plus.
🔎 In summary
🧠 Recommended for
- Computer vision researchers
- AI editing tool developers
- Interactive publishing projects
🔧 Compatible tools
- PyTorch
- TensorFlow
- Diffusers
- Image annotation tools
💡 Tip
To fully exploit, test multi-turn scenarios in a chain for realistic results.
Frequently Asked Questions
Does this dataset contain manual annotations?
Yes, all instructions and changes are annotated manually to ensure quality.
Can this dataset be used for multi-turn scenarios?
Yes, it includes single-turn and multi-turn image-editing scenarios.
Is this dataset suitable for training high resolution models?
The main dataset is in moderate resolution, but it is possible to adapt it for high resolution models via fine-tuning.
