20240037870. METHODS AND APPARATUS FOR DETERMINING AND USING CONTROLLABLE DIRECTIONS OF GAN SPACE simplified abstract (L'OREAL)
METHODS AND APPARATUS FOR DETERMINING AND USING CONTROLLABLE DIRECTIONS OF GAN SPACE
Organization Name
Inventor(s)
Parham Aarabi of Richmond Hill (CA)
METHODS AND APPARATUS FOR DETERMINING AND USING CONTROLLABLE DIRECTIONS OF GAN SPACE - A simplified explanation of the abstract
This abstract first appeared for US patent application 20240037870 titled 'METHODS AND APPARATUS FOR DETERMINING AND USING CONTROLLABLE DIRECTIONS OF GAN SPACE
Simplified Explanation
Methods, apparatus, and techniques in this patent application relate to determining directions in GAN latent space and obtaining disentangled controls over GAN output semantics. This enables the generation of synthesized images that can be used to train another model or create augmented reality experiences. The disclosed methods utilize the gradient directions of auxiliary networks to control semantics in GAN latent codes. It is demonstrated that even small amounts of labeled data (as few as 60 samples) can be used, which can be obtained quickly with human supervision. Additionally, important latent code channels can be selected with masks during manipulation, resulting in more disentangled controls.
- The patent application focuses on determining directions in GAN latent space and obtaining disentangled controls over GAN output semantics.
- The methods utilize the gradient directions of auxiliary networks to control semantics in GAN latent codes.
- Minimal amounts of labeled data (as small as 60 samples) can be used, which can be obtained quickly with human supervision.
- Important latent code channels can be selected with masks during manipulation, resulting in more disentangled controls.
Potential Applications
- Generating synthesized images for training other models.
- Creating augmented reality experiences.
Problems Solved
- Lack of control over GAN output semantics.
- Limited availability of labeled data for training GANs.
Benefits
- Enables precise control over GAN output semantics.
- Reduces the amount of labeled data required for training GANs.
- Provides more disentangled controls for manipulating GAN latent codes.
Original Abstract Submitted
methods, apparatus and techniques herein relates to determining directions in gan latent space and obtaining disentangled controls over gan output semantics, for example, to enable use of such to generating synthesized images such as for use to train another model or create an augmented reality the methods, apparatus and techniques herein, in accordance with embodiments, utilize the gradient directions of auxiliary networks to control semantics in gan latent codes. it is shown that minimal amounts of labelled data with sizes as small as 60 samples can be used, which data can be obtained quickly with human supervision. it is also shown herein, in accordance with embodiments, to select important latent code channels with masks during manipulation, resulting in more disentangled controls.
