[Feature Request] Add width x height ratio

[yurik44]

A small text telling the current aspect ratio would help a lot, I usually work with 16:9 and 21:9

[dfaker]

Do we mean displayed as a decimal? or the simplest ratio like you have up there?

[yurik44]

As I said it must be more difficult, but 16:9 = 1.777 and 21:9 is 2.3333

[EyeDeck]

Not a complicated equation anyway

def to_ratio(w, h):
    gcd = math.gcd(w, h)
    return (w//gcd, h//gcd)

>>> to_ratio(512,512)
(1, 1)
>>> to_ratio(1920,1080)
(16, 9)
>>> to_ratio(512,832)
(8, 13)

[yurik44]

That would be more than enough 🙂

[bmaltais]

This would be nice... a bit like what midjourney use to quickly set image size based on input like ar:1:2

Also this could be implemented with the fact that SD lose coherency past ~290k pixels and optionally propose resolutions that are more coherent. There is a discussion about that here:

https://www.reddit.com/r/StableDiffusion/comments/xhqada/please_help_confused_as_to_whats_going_wrong/

I did test the hypothesis and it is true... keeping the number of pixels under 290k will produce coherent output almost 100% of the time. Sticking with resolutions marked in green tend to produce the best outputs. You then feedback the produced output to img2img to scale it up and keep the coherency at a higher resolution while preserving the aspect ratio.

My guess is that the real coherency limit is 262144 pixels... matching the training pixel resolution of 512x512 images. Staying close to this maximum limit will produce mostly coherent images... but going well beyond will tend to create multiple arms, bad faces, duplicate subjects, loss of focus, etc, etc.

Those are an example of coherent resolution I calculated... they allow for slight overrun of 262144 pixels:

Light green: just under the optimal dark green pixels count
Dark green: equal on just under the 512x512 pixels count
Light yellow: just above the 512x512 pixels count. Should produce mostly coherent results
Light red: will start to produce more incoherency

Perhaps this is outside this PR request... but worth keeping in mind when generating images none the less.

[precompute]

That is an awesme metric. Did you generate it?

[KristerNielsen]

It would be fairly easy to add UI wise, when you drag the width/height sliders, to show in real time somewhere what aspect ratio it is. You could also add a few common aspect ratio lock buttons, and when you click them, the sliders moves in steps according to the locked aspect ratio.

[Wax-Replica]

i agree... i'd also like a button to lock aspect ratio to the present one, or lock aspect ratio to the closest possible to the input image for img2img.