Customized Diffusion Image Editing for Identity Preservation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image editing technologies using generative artificial intelligence fail to adequately preserve the original identity of images during editing.
Innovation Solution
A user terminal and server system utilizing a pre-trained artificial neural network with user-customized learning to edit images, involving primary and secondary user-customized learning processes to maintain image identity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a pre-trained diffusion model is used for image editing, then image generation capability is improved, but original identity preservation deteriorates
Solution Approach 1:
The system performs preliminary user-customized learning on the pre-trained diffusion model using the user's input image and related images before performing the actual image editing operation. This preliminary adaptation phase allows the model to learn user-specific features and preferences, thereby preserving original identity while maintaining the powerful image generation capabilities of the pre-trained model.
2Reliability
If user-customized learning is performed on a pre-trained neural network, then identity preservation is improved, but processing time increases
Solution Approach 1:
The system performs partial customized learning by selectively training only the necessary components of the neural network using a limited set of user-provided images and related images. This partial learning approach achieves sufficient identity preservation without requiring exhaustive training, thereby reducing processing time while maintaining reliable identity preservation.
3Measurement precision
If multiple related images are used for customized learning, then editing accuracy is improved, but system complexity increases
Solution Approach 1:
The system uses a single pre-trained diffusion model that serves multiple functions: it performs both the customized learning process and the actual image editing operation. By making the model universal and adaptable through learning from multiple related images, the system achieves high editing accuracy without requiring separate specialized components for each function, thereby managing system complexity effectively.
Data Source
AI summary
Disclosed are a user terminal, a server, and a method of operation capable of preserving and editing identity of an original image by performing user-customized learning on a pre-trained artificial neural network. The user terminal according to an embodiment includes an interface unit, a communication unit, and a control unit, in which the control unit may receive an original input image to be edited from a user and a text prompt including information about the editing, and transmit the received original input image to a server including a first artificial neural network trained to edit the input image according to the input text prompt, and receive, from the server, an output image for the original input image edited according to the text prompt by the first artificial neural network subjected to user-customized learning based on the original input image, and output the received output image through the interface unit.