What's the problem with the MP4s? The decoding can't take that much of your GPU's resources. And if it's a latency thing, you could start loading the next video before the previous finishes... Maybe grabbing a few of the first frames if there is some kind of difficulty in switching between the two.
In SFML image loading and editing operations are performed on the CPU (RAM), and drawing operations on the GPU (VRAM). Everything is done on the CPU if the computer doesn't have a graphics card.
FFMPEG's native operation is CPU-based, while MP4 decoding is done in software via h264 (CPU). The advantage of this approach is that it's cross-platform. Is possible change the decoder to use CUDA, but we lose the aforementioned advantage.
We can conclude that standard decoding would be done on the CPU. This decoding, performed while LLM is generating, is what I'm to avoiding. Although I would have to implement the version from mp4s to truly see the impact.
After the release of My Summer Love, I'll certainly test it. If it could be run directly from mp4, the size of the video game would be substantially reduced and more animations could be included.
Thank you very much for the observation.