Allow transparent interpolation animation generation

[NickAcPT]

Currently, when using the --generate-interpolation argument, the output GIF file will have a black background if the images have a transparent background.

[lucidrains]

ok done ad5d07a

[NickAcPT]

That code just converts the transparent images back to normal RGB.

Won't that cause the generated GIF images to continue having the black background?

[lucidrains]

@NickAcPT i don't think GIFs support transparency, which is why you were getting the black background, so I'm converting each frame to an RGB (without the transparency) before the GIF conversion

[NickAcPT]

GIFs do support transparency, but I feel like it isn't the best format to be used here.
Is there an easy way to allow users to specify the output format for the interpolation?

Proposed formats:

• GIF (default)
• Frame-By-Frame (exported directly, but outputs to a new folder instead of directly inside the results folder)
• APNG (animated PNG)
• WebP (webp is more modern than GIF and supports transparency)

In the case there is no way to implement the last two, I feel like the frame-by-frame should be implemented to make it so we can use external tools to make them.

[lucidrains]

@NickAcPT oh oops you are right

is your goal simply to have a white background? I think I can get it to work without having to reach for one of these less supported extensions

[NickAcPT]

@lucidrains

is your goal simply to have a white background?

Not really. My goal is basically to have a way to deal with the transparency of the images myself.
We can probably scrap both APNG and WebP.
Maybe the idea would be to keep things as they are, revert ad5d07a, and just implement the feature where the individual images are exported to a folder frame by frame.

[lucidrains]

@NickAcPT ok, let's do the frame by frame approach

[lucidrains]

@NickAcPT whatcha training on? any cool results?

[NickAcPT]

whatcha training on? any cool results?

Nothing special, really haha. Got bored of seeing the same skins people use in Minecraft (yes, that children game lmao). So, why not try this. Although my results haven't been that great because I've focused much on training the model.

ok, let's do the frame by frame approach

Thanks so much :D

[lucidrains]

@NickAcPT try the new augmentation feature! it should greatly improve results if you are working with a small dataset

[NickAcPT]

try the new augmentation feature! it should greatly improve results if you are working with a small dataset

Well, my dataset has some repeated files (which I'll delete before retraining), but in total I got near to 400k files. It'll just take forever to train since I don't own any good hardware and am instead using Google Colab (yes, shame on me)

[lucidrains]

@NickAcPT ohhh 400k is more than enough lol

[lucidrains]

@NickAcPT ok done e40eb6f

[NickAcPT]

Closing as it was implemented. Thanks again for your amazing work!