The new method is only using keyframes to generate the thumbnails, which don’t rely on other frames to be decoded to transform them to JPEGs (keyframes are self contained, only intra-coding used). The heavy CPU usage from the old method was primarily in the video decoding, not the JPEG encoding.
Based on my testing of generating the JPEGs with an Intel GPU (vaapi decode and mjpeg_vaapi encode), the new keyframe method is faster than using the GPU.
