IMAGE EDITING METHOD AND ELECTRONIC DEVICE FOR PERFORMING THE SAME

Smart overview of the Invention

An innovative method for image editing using electronic devices involves obtaining an image and an edit prompt, then generating an edited image through a diffusion model. This process applies varying strengths of image generation across different regions, based on a segmentation map. The segmentation map divides the image into multiple regions, allowing tailored editing for each area.

Technical Background

The method leverages generative AI technology, which synthesizes new data by learning from large datasets. While generative models can produce impressive results, challenges arise with processing speed when different strengths are individually applied to each image region. This invention addresses such challenges by optimizing the process.

Methodology

The technique involves determining different image generation strengths through hyperparameters that reflect both image and text conditions. These hyperparameters vary across regions, guided by a segmentation map which identifies object and non-object areas. The method also includes generating initial noise and refining it through a noise prediction process using classifier-free guidance (CFG).

Device Implementation

An electronic device equipped with this method comprises a communication interface, processor, and memory storing instructions. These instructions enable the device to acquire images and prompts, apply the diffusion model, and adjust generation strengths per region. The device uses CFG to merge conditional and unconditional predictions, enhancing the precision of noise prediction across regions.

Additional Features

The approach allows for combining additional input data into the diffusion model for enhanced noise prediction. The edited image reflects the edit prompt more in some regions than others, ensuring nuanced and context-aware edits. This flexibility makes the method suitable for diverse applications in image editing tasks.