from what i understand there is tradeoff for this approach bc u dont have access to raw data compression is only as good as the model that compresses the data a

from what i understand there is tradeoff for this approach bc u dont have access to raw data compression is only as good as the model that compresses the data and it can also be decompressed/inverted from gradient to reconstruct whatever the raw data is....