This release is based on neural networks. I trained two versions of models:
robustness model
More robustness against compression, photo editing etc. But less visually similar to original image.
(MSerial#: RAWMS10RE5EP3)
visual similar model
More similar to original image but providing less robustness.
(MSerial#: RAWMS100RE5EP9)
There're no plans to release model training code as organizing those code takes too many efforts.