Latent-to-Latent Mapping for Identity-Preserving Face Editing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional image editing systems fail to retain the personal identity of individuals in edited images, often resulting in unrecognizable faces due to unintended changes in attributes like skin color, hair color, or facial structure when modifying specific attributes.
Innovation Solution
A latent-to-latent mapping network is trained using a multi-task loss function to preserve face identity while allowing user-specified changes to attributes, converting latent and target attribute vectors into a hidden representation with fewer dimensions, enabling the generation of modified images that maintain the original identity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional image editing systems modify attributes using machine learning models, then attribute changes are achieved, but personal identity is lost and the person becomes unrecognizable
Solution Approach 1:
The patent segments the attribute modification process into separate controllable components. The latent space is divided into identity-preserving dimensions and attribute-modifiable dimensions, allowing independent control over which aspects change and which remain stable. This segmentation enables precise attribute editing while maintaining personal identity recognition.
Solution Approach 2:
The patent changes parameters in the latent space by transforming the latent vector through learned mappings. By adjusting specific parameters in the latent representation rather than directly manipulating image pixels, the system achieves attribute modifications while preserving the underlying identity features encoded in the latent space structure.
2Adaptability or versatility
If machine learning models compute high-dimensional feature vectors for image editing, then attribute changes are enabled, but unintended changes occur in skin color, hair color, or facial structure
Solution Approach 1:
The patent applies local quality by making different parts of the latent space serve different functions. Specific regions or dimensions of the latent vector are designated for controlling particular attributes (e.g., expression, pose) while other regions maintain identity characteristics. This localized control prevents unintended changes in unrelated attributes like skin color or facial structure when editing specific features.
Solution Approach 2:
The system uses feedback mechanisms during training to learn which latent space transformations preserve identity while achieving attribute changes. By incorporating identity preservation losses and attribute accuracy losses, the model receives feedback that guides it to find transformations that modify only the intended attributes without affecting other features, thereby improving precision.
Data Source
AI summary
Systems and methods for image processing are described. One or more embodiments of the present disclosure identify a latent vector representing an image of a face, identify a target attribute vector representing a target attribute for the image, generate a modified latent vector using a mapping network that converts the latent vector and the target attribute vector into a hidden representation having fewer dimensions than the latent vector, wherein the modified latent vector is generated based on the hidden representation, and generate a modified image based on the modified latent vector, wherein the modified image represents the face with the target attribute.