Expression Driving Model Real-Time Video Generation

The patent segments the facial driving process into two independent parts: a pre-trained expression driving model that processes driving information, and a facial feature map generation module that applies the driving information to a target facial image. This segmentation allows the computationally intensive pre-training to be separated from real-time inference, improving real-time performance while maintaining generation quality.

The patent applies preliminary action by pre-training the expression driving model offline using a large dataset of facial images and driving information. This pre-training phase completes all complex computational tasks beforehand, leaving only lightweight inference operations for real-time video generation. The pre-trained model captures complex facial expression relationships without requiring complex computation during actual video generation.

The patent trains the expression driving model in advance using large datasets containing diverse facial expressions and corresponding driving information. This preliminary training allows the model to learn complex gesture expression relationships and preserve accuracy during inference. The training phase completes all data processing and pattern learning beforehand, so that real-time generation only requires applying the pre-learned relationships without reprocessing the large datasets.

The patent creates a simplified copy of the complex facial driving process through the pre-trained model. Instead of recomputing complex facial expression relationships in real-time, the system uses the pre-trained model's learned representations to quickly generate driving information. This copying approach maintains gesture expression feature accuracy by preserving the learned relationships while significantly reducing computational complexity during video generation.