DanielRapp / cat-filter Public
submited by
Style Pass
This also gives you the opportunity to choose which bounding box and face landmarks model to use. For both the bounding box and landmark models you have an option of uint8 and uint16 and for the landmarks model you also have an f32 option. These trade between computation speed and accuracy, with uint8 being the fastest and f32 the most accurate, and uint16 laying somewhere in between.
Everything I've written is under MIT. But the models were trained using the nvidia cat stylegan which is under CC-BY-NC. This may fall under fair use or it might not (in which case only non commercial use is allowed), I'm not sure.
Read more github.com/D...
