search

pkhungurn / talking-head-anime-2-demo

Demo programs for the Talking Head Anime from a Single Image 2: More Expressive project.
http://pkhungurn.github.io/talking-head-anime-2/
MIT License
1.13k stars 155 forks source link

Image #3

Closed graphemecluster closed 3 years ago

graphemecluster commented 3 years ago

Thank you for letting me open this pull request. I don't mind if you commit again yourself in case you don't want my name to appear in your repository.

graphemecluster commented 3 years ago

I think you should look for another free (better open-source) iOS application doing the same job, since they are only pure ARKit parameters. Some users are not willing to pay money, you know. Edit: Oh I just found that I should comment this in the issue.

graphemecluster commented 3 years ago

I did think about all cases. I admit that I did this intentionally. How about if the user opens a file with an alpha channel but with no alpha at all (all values on alpha channel are 1)?

A better method would be to implement chroma key.

Anyway, I would add back the alpha channel checking. Sorry for any inconvenience caused.

graphemecluster commented 3 years ago

Sorry, I accidentally remove the background blending part when I change the method to convert_output_image_from_torch_to_numpy. Originally I am going to remove the blend_with_background method because Windows does support virtual camera streaming with alpha channel.

graphemecluster commented 3 years ago

Oh sorry, yesterday I altered it in a hurry and I didn't test it at all. But actually I assumed you to edit them and commit to my branch before merging.

dragonmeteor commented 3 years ago

Let me take this opportunity to answer some questions and comments:

I think you should look for another free (better open-source) iOS application doing the same job, since they are only pure ARKit parameters. Some users are not willing to pay money, you know.

Thank you for the suggestion, but I simply have no motivation to do this right now.

How about if the user opens a file with an alpha channel but with no alpha at all (all values on alpha channel are 1)?

There's already a specification on how to prepare input images written in the README file. In this case, the image does not meet the specification, so the output is undefined. It is hard to check whether all the pixels have "correct" alpha values. However, it is much easier to check whether the image has an alpha channel at all. So, I think it is a good idea to tell the user of the mistake when it is easy to do so.

A better method would be to implement chroma key.

To keep the features of this demo minimal, I currently do not plan to implement chroma key or include it as a feature.

But actually I assumed you to edit them and commit to my branch before merging.

I believe this is your pull request. It would not be respectful to you if I were to interfere with it.

I think the pull request is close to being done. Just only a few more bugs to fix. Please bear with me for a little bit more.

graphemecluster commented 3 years ago

Actually that's because I totally forgot about the notebooks – It works great after alternations. Some lines changed while exporting the notebooks (because the .ipynb format was updated?), but I changed back a little bit to conform with yours.

Other minor changes include:

I believe this is your pull request. It would not be respectful to you if I were to interfere with it.

I would uncheck the Allow edits by maintainers box if I don't want others to touch my branch.

dragonmeteor commented 3 years ago

Please push. Thank you for your hard work.

graphemecluster commented 3 years ago

No, I should be thanking you. … Um, I don't think I have write access to this repository.

pkhungurn commented 3 years ago

Ah, sorry. I've merged it myself.

graphemecluster commented 3 years ago

I forgot to also remove the FAQ part from README.md. And one more thing: Just for curiosity, why you separated the model files to Dropbox? LFS might be helpful if the files are too large.