The application discloses a desktop cloud scene-based AI
large model traffic optimization method, and belongs to the technical field of desktop
cloud computing. The method comprises the following steps: a
server end performs frame-by-frame screen acquisition on a
virtual desktop at a
frame rate to output an original
image frame; interframe difference calculation is performed on the original
image frame and a previous frame cache image to output a difference pixel proportion; the original
image frame is input into a visual
language model to perform semantic reasoning and output a pixel-level importance
heat map; region extraction is performed on the pixel-level importance
heat map to obtain a rectangular region and an average importance
score; rectangular region data extracted from the previous frame is input into an
encoder parameter configuration interface; the
encoder queries a preconfigured QP offset mapping table and writes into a QP offset register to perform
differential coding and output an encoded code
stream. The application realizes
differential coding based on region importance, adaptive reasoning frequency switching and asynchronous pipeline
parallel processing, and improves the visual quality and
bandwidth utilization efficiency of desktop cloud transmission.