search

k-l-lambda / stylegan-web

A web porting for NVlabs' StyleGAN.
160 stars 41 forks source link

Feature suggestion - merger animation exporting #5

Open leoffx opened 3 years ago

leoffx commented 3 years ago

It would be very useful to have the ability to export the intermediate "Merge" images into a gif.

Also as a major feature suggestion:

k-l-lambda commented 3 years ago

It would be very useful to have the ability to export the intermediate "Merge" images into a gif.

Interesting. And there are a lot of interpolating handlers, how do you want to animate them? Or maybe by a group of parameters, how to define the configuration?

Also as a major feature suggestion:

  • having the ability of learning and exploring meaningful latent directions

How to do this? Tell some details about your ideas.

leoffx commented 3 years ago

Interesting. And there are a lot of interpolating handlers, how do you want to animate them? Or maybe by a group of parameters, how to define the configuration?

Well, at first I was imagining just sampling the intervals from the "ALL" slider, and merging the results into a gif, like it's done with projection animation.

How to do this? Tell some details about your ideas.

I'll do more research on this, but I've seen some papers such as Unsupervised Discovery of Interpretable Directions in the GAN Latent Space , which I think goes in this direction.

k-l-lambda commented 3 years ago

I'll do more research on this, but I've seen some papers such as Unsupervised Discovery of Interpretable Directions in the GAN Latent Space , which I think goes in this direction.

I think you can get some inspirations by latent vectors linear computation experiment in this web page. However, a deep research require a lot of computation, web is too inefficient.

Well, at first I was imagining just sampling the intervals from the "ALL" slider, and merging the results into a gif, like it's done with projection animation.

Yes, we can make a good video composing tool, I'll consider this.