Add simple variant mechanism

[Oceanswave]

This pr adds the following:

• Introduces a new flag, [-v, --variants] that accepts an int and pipes generated images back through the img2img pipeline in order to have a simple 'variants' method. This obvi has some limitations: works solely on strength, and uses the original prompt, but provides some interesting results
• Fixes the warning that LANCZOS is deprecated introduced by a different PR
• Hopefully better resolves the device issue for folks running on different device types in the ddim and plms samplers

[morganavr]

This is a prompt I provided:

dream> "photo dog labrador park" --init_img=./init-images/dog.png --strength=0.75 -s50 -n1 -v 5

Created variants images are very different from each other. If I remove "-v 5" parameter and run this prompt instead:

dream> "photo dog labrador park" --init_img=./init-images/dog.png --strength=0.75 -s50 -n 5

then script creates images of the same quality. Is it an expected behavior of this PR?

Below are photos from your PR with "-v 5" argument:

[lstein]

Thanks for this. Is there a reason that you created an unused variants argument in the calls to img2img() and text2img()? Is this just to avoid python complaining about an unexpected argument?

[Oceanswave]

@morganavr the strength of how closely img2img follows the image is controlled by the -f or --strength flag. 1 is completely different, 0 is completely the same. the default is 0.75 which is fairly different, but you'll notice that certain features are kept in the same place, such as the nose, eyes. tweaking the strength to -f 0.5 or lower makes it closer and closer to the original image

@lstein sheepishly yes - open to a better way, delete the property prior to pass?

[lstein]

@Oceanswave I'm inclined to leave it be. I just got a little confused because I was looking for the variation-generating code in the simplet2i (which ain't simple any longer) and thought you'd forgotten something. The code is seriously in need of refactoring anyway, and I'm likely going to be changing how arguments are passed to the two routines.

[morganavr]

@Oceanswave
my concern with that PR is that even without --variants argument script produces the images of the same variation similarity.
I thought that --variants will produce almost identical images that differ only slightly from each other.

Can you provide a usage example with correct parameters?
Also I think that README should be updated because how users are supposed to know about this new feature? :)

[Oceanswave]

@morganavr Sure -

Snap a dog wearing swim goggles and a funny hat at the park -n1 -f0.4 -v5

[Oceanswave]

@morganavr since this is pretty fun.. here's another

Snap a rocket ship made of cheese -n1 -f0.6 -v5

[Oceanswave]

@morganavr

Snap pancakes that have the outline of a gorilla -n1 -f0.6 -v5

burning the pancake to get the effect is... pretty clever..

I like the smiling one with the hand

[Oceanswave]

@morganavr

Snap a butterfly with stained glass wings -n1 -f0.6 -v5

[Oceanswave]

okok, one more.

A painting of a beach at night, stars flicker and reflect in the waves, in the style of Thomas Kinkade, ultra realistic -n1 -f0.6 -C 13 -v5

[Oceanswave]

Reducing the -f or --strength reduces the variation

Snap a man wearing a pineapple hat. -n1 -f0.2 -C 13 -v5

Huh.. I just thought of another feature... :)

[Oceanswave]

And it's possible to supply your own initial image as well -- just using one that is in the repo

a photo of a dog sitting on a bench -Idata/inpainting_examples/overture-creations-5sI6fQgYIuo.png -f0.2 -C17.0 -v 5

[vekst]

This'll be fun to use, nice one. Somewhat related, but this gives me an idea for something similar but for normal, non img2img generations. A variant command that changes the CFG Scale of a generation which is using a seed. Would be kinda fun.

edit: disregard, just noticed something similars been thought of in the Issues of the main project.

[morganavr]

Thanks for all these examples @Oceanswave
It seems that I used wrong parameters and your feature works as intended!