Neural Face Morphing with Blended Identity Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer-generated animation and facial image manipulation technologies struggle to smoothly and temporally consistently morph faces between identities while preserving certain aspects of the original identity, such as gender or ethnicity, without perceptible artifacts.
Innovation Solution
A face-morphing model utilizing a shared set of trainable neural-network parameters and identity-specific parameters is trained to blend characteristics of multiple input identities, employing an autoencoder with a shared encoder and identity-specific decoders, and a mask decoder to generate a blended face image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If face morphing is performed between multiple identities, then the ability to transform facial characteristics is improved, but temporal consistency and smoothness deteriorate due to perceptible artifacts
Solution Approach 1:
The patent segments the face morphing process into distinct components: an autoencoder for feature extraction, identity-specific parameter sets for different facial characteristics, and a blending mechanism using interpolation parameters. This segmentation allows each component to be optimized independently, improving temporal consistency while maintaining transformation capability.
Solution Approach 2:
The patent employs parameter changes by using identity-specific neural network parameters that can be blended through interpolation. By adjusting interpolation parameters, the system smoothly transitions between different identity characteristics, ensuring temporal consistency while maintaining the ability to transform facial features.
2Adaptability or versatility
If identity-specific parameters are used for each input entity, then the ability to preserve specific identity characteristics is improved, but device complexity increases
Solution Approach 1:
The patent implements universality by designing a shared autoencoder structure that processes all input identities, while only the decoder parameters are identity-specific. This multi-functional design allows the same base architecture to handle multiple identities, reducing overall complexity compared to having completely separate models for each identity.
Solution Approach 2:
The patent applies the nested doll principle by embedding identity-specific parameters within a hierarchical structure: the autoencoder provides a shared feature extraction layer, while identity-specific decoders are nested within this framework. This nesting allows compact representation of multiple identities without proportionally increasing complexity.
3Manufacturing precision
If smooth face morphing is achieved, then visual quality is improved, but computational requirements increase to maintain temporal consistency
Solution Approach 1:
The patent applies preliminary action by pre-training the autoencoder on diverse facial data before use. This pre-processing step establishes a robust feature extraction foundation that reduces computational burden during actual morphing operations, allowing smooth transitions with lower real-time computational requirements.
Solution Approach 2:
The patent uses copying by leveraging the shared autoencoder structure that can be reused across different identity transformations. Instead of computing unique transformations for each identity pair, the system copies and adapts the learned feature extraction process, reducing redundant computations while maintaining high visual quality.
Data Source
AI summary
A method morphs an input image depicting a face to an output image depicting a face that is a blend of characteristics of a plurality of input entities. The method comprises: training a face-morphing model comprising: a shared set of parameters shared between the input identities; and, for each of the input entities, an identity-specific set of parameters. The method also comprises: receiving an input image depicting a face of one of the plurality of input identities; receiving a set of interpolation parameters; combining the identity-specific sets of trained neural-network parameters for the plurality input identities based on the interpolation parameters, to thereby obtain a blended set of neural-network parameters; and inferring an output image depicting a face that is a blend of characteristics of the input entities using the shared set of trained neural-network parameters, the blended set of neural-network parameters and the input image.