Add image-to-image and image-generation tasks

[nateraw]

Adds 2 new tasks to filter by:

• image-to-image
• image-generation

Adding these as we will encourage folks to add these tags in the HugGAN event.

⚠️ When completing this PR, I noticed lib/ was blocked in the .gitignore, so I removed it. Otherwise I would have to force add files. Hopefully that's ok 😅

[osanseviero]

Thanks! This is cool

• Do you have a screenshot with this built just to confirm it works and how it looks like?
• I was wondering if image-generation should be called unconditional-image-generation to clarify that there is no input, but I was also wondering if I might be missing something

[osanseviero]

cc @julien-c

[nateraw]

As requested, here's a screenshot

[osanseviero]

Thank you! Overall looks good and I think the icons are great. My only remaining question is if the task should be image-generation or unconditional-image-generation (for the cases in which the image is generated from noise)

[julien-c]

not super familiar with this task, but what's the input for image-generation? just a seed?

[osanseviero]

Usually it's a noise vector (the seed is used to have consistent noise vector for consistent output)

[julien-c]

unconditional-image-generation sounds weird/vague to me but I don't have a better idea right now =)

[julien-c]

noise-to-image?

[osanseviero]

No strong opinion from my side, I think noise-to-image could work as well

[julien-c]

ok let's do this unless @nateraw objects then!

[nateraw]

@julien-c I think noise-to-image is actually more confusing...unconditional-image-generation would actually fit Papers with Code's taxonomy best...my only objection to that is that the name is a little long haha.

Basically image-generation is the parent task and unconditional-image-generation and conditional-image-generation fall within it. So I vote we either stick with image-generation (because I doubt we'll end up enabling inference API for the sub-tasks since the models vary so greatly in size of the noise vectors, how conditions are chosen, etc) or we go with unconditional-image-generation to match the taxonomy on PapersWithCode.

WDYT @osanseviero

[NielsRogge]

Yeah it's actually only GANs that start from noise for unconditional image generation, right? Not sure how diffusion models and VAEs do this. noise-to-image is probably too GAN-specific

[julien-c]

the drawback is that for me image-generation is very easily confused with text-to-image

(Note: i don't think we want to model PWC's taxonomy too much, since ours aims to be simpler)

[julien-c]

with this in mind, maybe unconditional-image-generation is the best one then yes =)

[nateraw]

Done. Screenshot here for your viewing pleasure.

[osanseviero]

Nice, thanks for this! 🚀