METHOD AND DEVICE WITH PERSONALIZED IMAGE GENERATION

Smart overview of the Invention

A novel approach for personalized image generation involves creating a user-specific preference vector. This vector is derived from several preset properties, which guide the customization process. The system then generates a personalized image by integrating this preference vector with an image generative model and various characterization layers. These layers correspond to the predefined properties, ensuring the output aligns with the user's unique preferences. The final image is produced in response to an image generation command.

Technical Background

The technology leverages advancements in artificial intelligence, particularly in text-to-image generative models. Traditionally, image generation based on user preferences involved building datasets and training deep learning models with reinforcement learning or direct policy optimization. This invention refines the process by incorporating user preference data directly into the model through characterization layers, enhancing the alignment of generated images with individual user tastes.

Process Details

The system's core functionality lies in its ability to modify the image generation process dynamically. It applies input data to characterization layers connected in parallel to the model's layers. The weight of each characterization layer is adjusted according to the user's preference vector, allowing for a tailored aggregation of outputs. This dynamic adjustment ensures that the final image reflects the user's desired properties, such as brightness, saturation, contrast, and edge sharpness.

Preference Vector Generation

To create the preference vector, users are presented with multiple images, each generated with varying weights across the characterization layers. Users select their preferred image, which updates the preference vector, refining future image outputs. This iterative process helps the system learn and adapt to individual preferences over time, enhancing the personalization of generated images.

Training and Implementation

Training involves developing a preference prediction model using a dataset that includes images and their property scores. Characterization layers are trained to maximize specific property scores, such as brightness or contrast, while keeping the generative model's parameters constant. This method ensures that the layers enhance the image properties most valued by users. The technology can be implemented on devices with processors capable of executing the described methods, supporting both text and voice input for image generation commands.