search

bryandlee / animegan2-pytorch

PyTorch implementation of AnimeGANv2
MIT License
4.38k stars 640 forks source link

mypfp.io, an anime-like NFT avatar generator service #25

Open dressedanonymous opened 2 years ago

dressedanonymous commented 2 years ago

Thanks for publishing a very useful model and sample code. We've been using your "Face Portrait v2 model" in our experimental NFT project mypfp.io, an anime-like NFT avatar generator service. Since it was implemented in PyTorch, we were able to quickly develop an API using Flask.

We've put a sample video on mypfp.io for you to play with if you want and we would be happy if you could add some collaboration projects section on the README.md.

Thanks!

mypfp.io NFT minted in the video

https://user-images.githubusercontent.com/94442111/141962737-b5d1ad1f-9729-447f-b9fa-c82ce5b27972.mp4

Kitty-sunray commented 2 years ago

Thanks for publishing a very useful model and sample code. We've been using your "Face Portrait v2 model" in our experimental NFT project mypfp.io, an anime-like NFT avatar generator service. Since it was implemented in PyTorch, we were able to quickly develop an API using Flask.

We've put a sample video on mypfp.io for you to play with if you want and we would be happy if you could add some collaboration projects section on the README.md.

Thanks!

mypfp.io NFT minted in the video mypfp_short.mp4

You sure license allows that?

dressedanonymous commented 2 years ago

MIT License

Kitty-sunray commented 2 years ago

there is no training code available in this repo -> you didn't train anything yourself and just used weights trained on dataset you didn't own or even saw, and most likely is original animegan2 dataset. original animegan2 and it's dataset is "for non-commercial purposes such as academic research, teaching, scientific publications." as written on project page.

dressedanonymous commented 2 years ago

About Face Portrait v2 model, it says they used training sets prepared by themselves to derive weights and publish its weight and model under the MIT license. It is not related to the original project.

Kitty-sunray commented 2 years ago

there is no proof of their words since dataset and train code are not provided, and results (weights) cant be replicated.

dressedanonymous commented 2 years ago

Have you ever checked original datasets? Anyone who has studied machine learning even a little bit can easily imagine how difficult it would be to create such a portrait-specific model from the original data set.

Kitty-sunray commented 2 years ago

ha, have YOU seen datasets weights you used was trained on? was able to reproduce the weights with ~same result thus claims of author are believable true? So far it looks like model is trained on random googled images/movies with zero respect to rights, and author claims "all ok it is mit" while providing zero proof

dressedanonymous commented 2 years ago

If you feel that way, then why don't you just not use it yourself?

EuphoricPenguin commented 2 years ago

I'm actually curious what the authors of this repository intended with the license. My non-legal interpretation is that the MIT license would include the weights, but I'd like to here from the repo owner themselves.

My understanding is that the copyright and licensing associated with a dataset need not necessarily apply to the model and consecutive generated outputs. Working with separate licenses for the code, datasets, weights and outputs can be incredibly confusing, and it's fair to say that it's up in the air.

Georgexp97 commented 7 months ago

This could be really interesting. I think such someone will be interested in NFT Avatars. If someone is looking for an answer whether it is worth investing in nft avatars, I highly recommend this post: https://gamerseo.com/blog/are-nft-avatars-worth-the-hype-around-them/