[Feature Request]: Symmetrical Outpainting
Closed this issue · 2 comments
Is your feature request related to a problem? Please describe.
Images of the problem I'm running into:
https://www.reddit.com/r/StableDiffusion/comments/10jprya/how_do_i_get_symmetrical_outpainting/
Describe the solution you'd like
A checkbox that allows the outpainted area to be symmetrical.
Proposed workflow
-
User can select the symmetrical check box
-
If checked, expanded area matches on both sides
** Might require Left/Right or Up/Down only -
Flow for expanding an image left and right:
-
256px of space added to the left side of the canvas
-
X steps happen
** suggested 2-4 steps -
Copy the 256x512 image
-
Add 256px to the right of the canvas
-
Flip and paste in the copied image
-
Repeat until done
-
Optionally allow X steps at the end to remain on respective side
** Allows for unique features to appear
Describe alternatives you've considered
Currently, I'd have to photoshop one side onto another.
Additional context
Also posted here:
AUTOMATIC1111/stable-diffusion-webui#7147
hey, so i don't want to immediately crush anyone's hopes or anything but this seems like an incredibly niche desire, like genuinely something i've literally never heard anyone even mention doing before, and seems to be only applicable to extremely limited use cases with very specifically chosen input images... like unless both edges are essentially 10-15% of the image's width of a flat single color, there's no possible way this could ever produce coherent output without manual intervention in the first place.
it really sounds like photoshop is a far more appropriate tool for the job, or possibly just an A1111 script?
[edit] ok wow so i just re-read your request and unfortunately i have to doubly drop some bad news- i don't know of anything in A1111 webUI that supports pausing/resuming image diffusion from latent noise in the first place, much less doing something inbetween the pause/resume, so there'd be nothing we'd even be able to do here in the first place
i think this is exactly what you're looking for as far as mid-diffusion processing? comfyanonymous/ComfyUI@69df7eb ... comfyanonymous/ComfyUI@3fa009f too maybe to flip the latent?