US20260320974
2026-10-15
Physics
G06V10/7715
The patent application addresses challenges in artificial intelligence (AI) by introducing a multiple-input, fixed-size representation for generative tasks like image generation. This representation preserves identity features from multiple inputs, enabling the creation of high-quality images of subjects in various poses and environments. The approach involves training an encoder with a multi-decoder strategy, allowing the representation to be utilized by different decoders for diverse objectives.
The disclosure pertains to digital imaging systems, focusing on camera systems and image processing hardware for capturing and generating identity-preserving representations. It aims to enhance machine learning models used in complex AI applications, addressing issues of computational efficiency and output quality.
Machine learning models face challenges in preserving identity features when generating new images from existing ones. The disclosed method encodes multiple images into a single, fixed-size representation, maintaining a consistent format for downstream applications. This representation captures detailed identity information by combining visual data from various inputs, improving accuracy and fidelity in generative tasks compared to single-input methods.
An encoder is trained to generate representations for generative tasks, using image features extracted from a set of images. The encoder employs a transformer-based architecture with cross-attention and self-attention layers. Training involves a few-to-many reconstruction objective with multiple decoders, optimizing the encoder to minimize loss between generated images and the original set.
The trained encoder generates representations for downstream tasks like image generation, using images sampled from videos. Recognition algorithms identify subjects in these images, which are clustered by pose and selected based on quality. The resulting representations facilitate high-fidelity, identity-preserving outputs, reducing overfitting and supporting generative objectives. This structured approach consistently produces superior results in generative tasks.