stablediffusion-infinity icon indicating copy to clipboard operation
stablediffusion-infinity copied to clipboard

Published 20 hours ago • verified publisher icon lkwq007

Bad result

Open egorpezdir opened this issue 2 years ago • 5 comments

I am running a local notebook and am getting very bad false results. Example Example

egorpezdir avatar Sep 16 '22 13:09 egorpezdir

outpaint_20220916

It also requires prompt engineering. (here I use "Ivan Grohar, impressionist, painting style") For high resolution image, you may need to resize the image or increase the size of selection_box for outpainting

lkwq007 avatar Sep 16 '22 15:09 lkwq007

It also requires prompt engineering. (here I use "Ivan Grohar, impressionist, painting style") For high resolution image, you may need to resize the image or increase the size of selection_box for outpainting

Oh so the program itself can't pick up the style that good without specifying it?

egorpezdir avatar Sep 16 '22 15:09 egorpezdir

It also requires prompt engineering. (here I use "Ivan Grohar, impressionist, painting style") For high resolution image, you may need to resize the image or increase the size of selection_box for outpainting

Oh so the program itself can't pick up the style that good without specifying it?

I think the current version of diffusers's stablediffusion pipeline is unable to do that. (Currently there is no explicit constraint to make the filled region looks like other parts.) My program only provides a handy UI and the initial guess for empty region based on texture synthesis.

lkwq007 avatar Sep 16 '22 15:09 lkwq007

It also requires prompt engineering. (here I use "Ivan Grohar, impressionist, painting style") For high resolution image, you may need to resize the image or increase the size of selection_box for outpainting

Oh so the program itself can't pick up the style that good without specifying it?

I think the current version of diffusers's stablediffusion pipeline is unable to do that. (Currently there is no explicit constraint to make the filled region looks like other parts.) My program only provides a handy UI and the initial guess for empty region based on texture synthesis.

Understandable. I just saw your demo and thought it worked that way.

egorpezdir avatar Sep 16 '22 16:09 egorpezdir

@lkwq007 I am thinking through how some of the current outpainting techniques work. Could this use img2img and treat this as inpainting, using the previous results as a mask for the next image? (If that's already how this works, forgive me).

I've seen pretty good results using the newwer outpainting custom script in AUTOMATIC1111's repo, but the UI for it is pretty inflexible. Link to the script: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/84e97a98c5233119d0f444e0a3a0f6391da23677/scripts/outpainting_mk_2.py

Can any of this be used to make these results better?

JamesIV4 avatar Oct 01 '22 19:10 JamesIV4