The LoRa was trained on just two single transformer blocks with a rank of 32, which allows for such a small file size to be achieved without any loss of quality.
Since the LoRa is applied to only two blocks, it is less prone to bleeding effects. Many thanks to 42Lux for their support.