Magic Insert
Drag and drop image editing for style perception
- Style aware personalization: By training LoRA delta and text tagging, a personalized diffusion model is used to match the target image style.
- Object Insertion: Using Bootstrap Domain Adaptation technology, realistic objects from specific domains are inserted into the model to adapt to diverse artistic styles.
- LLM guided pose modification: using Large Language Model (LM) guided pose adjustment to provide reasonable pose and environment interaction for image regions.
- Bootstrap domain adaptation results: Adapt the effective domain of the model through self output subsets, improving the ability to process stylized images.
- Comparison of style perception personalized baseline: Compared with existing technologies, Magic Insert has a significant advantage in overall quality.
- Attribute modification: allows modification of key attributes of the theme, such as character reshaping or adding accessories, providing flexibility for creative use.
- Editability/fidelity trade-off: demonstrates the trade-off between fidelity and editability of the theme in different fine-tuning iterations.
Product Details
Magic Insert is an innovative image editing technology that allows users to drag and drop any style of image theme onto another style of target image, achieving style perception and realistic insertion. This technology formally defines the problem of style aware drag and drop by solving two sub problems of style aware personalization and real object insertion in stylized images, and proposes a method to solve it. The method of Magic Insert is significantly superior to traditional image restoration techniques. In addition, a dataset called subjectPlop is provided to facilitate evaluation and future development in this field.
