Unfortunately, my current rig can’t handle SD, so I was happy to stumble across Perchance.org
I am familiar with prompt writing and SD functionality due to using other web based generators…
Some of the features I miss are shown in the attached. Obviously I wouldn’t expect all of these to be implemented here, but perhaps this could be used for some ideas?
These 3 are definitely handy:
The cfg (guidanceScale) is a LOT more sensitive (less granular?) here. On that site, I regularly went as high as 12 with excellent results. Here, things start to get pretty bad above 8 or 9.
gfp (GFPGAN) works excellently. Only a setting of 1 or 2 is used. I notice here, that if you generate a full body pic (or not even that sometimes), faces get pretty melty.
The Variant Seed thing is pretty neat - it lets you use a percentage of a seed image.
Whatcha think @[email protected] ?
You must log in or register to comment.
I think the gfp thing will become unnecessary soon so it’s probably not worth me adding, but in the meantime you can download the image and use the gpf model separately e.g. here: https://replicate.com/tencentarc/gfpgan
For the “variant seed”, I think this is basically “image-to-image”? I.e. choose a starting image and a prompt, and choose the “strength” of the image (i.e. amount of random noise added to it). So basically it’s like adding an image prompt alongside the normal text prompt. If that’s what you’re referring to here, then yep Perchance will have that within the next few weeks hopefully. Currently waiting on a bug to be fixed in ‘upstream’ code that I’m using.
Unfortunately gfgpan is Nvidia only. I’m still using a Radeon card. =/
The link I shared is an online tool, so it doesn’t matter what your local GPU is
Ah thanks… I didn’t see the login… thought it was just a github project for download.
Controlling the steps/iters might be nice but I know higher steps/iters will hog the server since it will take longer time to generate the image.
