New feature: Export images

[StableLlama]

This PR adds the ability to export the images to a new folder. During this export the images can be resized to suit the target model size and cropped to fit into the target buckets. After the resize a slight sharpening is happening as it is recommended by graphic specialists.

Notable features:

• Resizing is done with the highest quality algorithm
• Best target dimensions are thoroughly searched
• Color space conversion can be applied as source images might use different color spaces but the training tools usually don't know anything about color spaces
• A statistical preview is shown

Additionally code changes:

• The use of the DEFAULT_SETTINGS mechanism was simplified to reduce the necessary coding to prevent inconsistency bugs and improve readability

[jhc13]

I appreciate your work, but I think this might be outside the scope of the program. TagGUI is designed for creating and editing text captions for images without modifying the images themselves. I think the features you are implementing are better suited for a dedicated program for processing the images.

By the way, TagGUI already supports moving or copying images to a different directory, but these actions do not alter the images.

[StableLlama]

This PR is the base for the next two steps I want to include as well:

1. support manual cropping
   • Give the user a hint to catch exactly relevant aspect ratios (due to the bucketing it's not an obvious task)
   • Give the user a hint about what will be additionally cropped to fit a bucket to be able to fine adjust the crop
   • Possibly: allow multiple crops per image, like having a full body crop and a face crop out of a high res source image.
2. support masking
   • Have one or more positive (rectangular) masks
   • Have one or more negative (rectangular) masks
   • When exporting create the real mask from the union/intersection of these masks and either store it in the alpha channel of the exported image or store it in a parallel directory, just like the usual training scripts are expecting them
   • Likely: show what the mask will exactly cover or not as the quantization due to the VAE / latent space might surprise the unaware user
   • Probably: have (rectangular) "hints" that show major features ("head", "hand"). Likely these can then help to quickly create a mask or crop, e.g. to prevent cropping though a hand which will create bad training results
   • Likely: The data storage for these features will be so simple that external tools can easily create them to allow a quick semiautomatic workflow. E.g. have a watermark detector tool to create a negative mask for it. Have a hand and face detector tool to create hints. The user can then quickly create the crop is such a way that all these information is taken care of. (Of course this functionality could be added to taggui as well, when that is desired and someone knowledgeable of YOLO models or similar can create the code)
   • Out of scope: a "real" mask editor where you can paint a mask.

All of this is the next step in the workflow of starting with a bunch of images and finally coming up with a ready set of training images. So far I know of no other tool that can do this, I know people who are missing a tool for this (not only me :)), it is closely related to tagging the images and taggui is the perfect base for this extended functionality.
So I think it is in the scope of taggui. Or, to be more precise: it is a valuable extension of the scope in an area where no tools are available.

And all of this is in the (highly appreciated!) spirit of not modifying the source images. That's even the reason for creating the exporting function as it allows you to reuse your valuable data set for new models when they become available:
You are using SD1.5 right now? Tag the images and export them with 512px.
Now you are switching to SDXL or Flux? Only thing to do is to do a new export with 1024px.
Then the next model comes that does UHD 4k? Still, only one more export, this time with 4k. (You can do it with the code from this PR already, even when it doesn't know about 4k)
Or a future model that does HDR and a wide color gamut? And your image collection already has such images? Then it's again only one export, just select the color space accordingly. (And all exports prior to this were even working correctly for you as it changed the color space to sRGB as that is what the current models are expecting. If you would have given those models these HDR images by plain copying the training result could have been disappointing)

[StableLlama]

Hint about HDR, the Open Model Initiative seems to look actively at supporting HDR images: https://github.com/Open-Model-Initiative/HDR_SDXL

[StableLlama]

@yggdrasil75 I've seen that you are working on real JPEG XL support. Do you intend to make a pull request out of it soon, as it would work very well with my latest commit to this PR (-> #337 )?

[yggdrasil75]

I will once I stop being lazy and actually have it working. might be tomorrow now that someone else mentioned it.

[yggdrasil75]

made the jxl pr. btw, the biggest benefit I would see with this would be adding segm/bbox support to auto mask non-character portions, or auto crop to character, is that the eventual plan?

[StableLlama]

That's exactly where I'm heading at.
But before the automation can come (when I actually have no experience at) the manual editing interface must work as you'll need that for refinement anyway.
And once that is here you can easily add automation for it.

[StableLlama]

Now I've also added a few more refinements, so I will not change this branch / RP anymore (unless someone finds bugs that need fixing, of course).
For the next step of the exporting (i.e. the crop editor) I'll create a new PR that will require this PR as a basis.

So please pull it, as well as #335, as both form the basis.
Also #337 would fit very well here.