>>105818371
It drastically reduces the memory requirements per frame. How well a model works with longer context/attention who knows, it'll likely need a Lora or finetune but I believe this means you can apply this to training as well which means we can train on longer samples.