[Request] Finally, Outpainting (like Dally) for SD!

[ZeroCool22]

trafficstars

https://github.com/lkwq007/stablediffusion-infinity

Could you include it in your UI?

https://user-images.githubusercontent.com/13344308/198017867-ff2a65cd-8406-4910-8d5f-0d8272b7e841.mp4

[aniketgore]

this looks awesome.

[dingir8]

Please ! <3 Automatic are so wounderfull man for all you already did for us, pls add this its very cool

[kodxana]

+1 for this I'm tired of having setup 5 containers for each web ui :D

[1blackbar]

Im gonna be nitpicky and say it does not convey a styule of the painting , maybe you need to lead with prompt... but yeah it does look good but i think this is using such different UI it would require major rewrite to this one to be able to move rectangle like this

[marquilio7]

Im gonna be nitpicky and say it does not convey a styule of the painting , maybe you need to lead with prompt... but yeah it does look good but i think this is using such different UI it would require major rewrite to this one to be able to move rectangle like this

You could add a new empty tab in the UI for this

EDIT: I would make the whole canvas and interaction on the client-side

[randecse]

This would be brilliant! Thanks for everything, Automatic.

[johneliot1978]

yes please!

[gamerey43]

i second this

it will help a lot

[gamerey43]

this would actually simplify the code for outpainting

Automatic this is literally a good idea and it's good the community views your version as the defacto for the most part :)

[MrFelt]

Yeah I would die for this ngl. Would make it easy as heck to outpaint.

[IkariDevGIT]

pleeeeeease

[nulqwy]

What about a collab with author of sd infinity?

[IIIIIIIllllllllIIIII]

please please please please please :)

[Mikerhinos]

This would be a great addition to me.

[mbloflin]

I work with very wide images (5:1) and SD does not handle this width well for the most part. With Infinity as part of the webui I could create better images and outpaint the extra width and have a more coherent subject. That would be awesome! Thanks for all that we have so far it has been GREAT.

[ClashSAN]

try the script basic outpainting canvas test or the PR #3693

[mezotaken]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Extensions#openoutpaint-extension Reopen if that's not the feature you were requesting for, or it doesn't work as stated.

[Frief84]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Extensions#openoutpaint-extension this is something different! ;-)

[mezotaken]

@Frief84 Can you elaborate a bit? I can see its missing eraser, and it is obviously different UI, but everything else is in place? Regional outpainting on infinite canvas using local webui installation. It's no problem to reopen the issue, but we need to point out a difference so that someone else who sees the issue could understand this difference and implement it.

[Frief84]

@Frief84 Can you elaborate a bit? I can see its missing eraser, and it is obviously different UI, but everything else is in place? Regional outpainting on infinite canvas using local webui installation. It's no problem to reopen the issue, but we need to point out a difference so that someone else who sees the issue could understand this difference and implement it.

oh yes of course sorry An eraser is only present...but just on different place

So the first difference I found is the infinite canvas, in the implemented openoutpaint extension it is limited, which means you can create larger resolutions in this one

The second is that they seem to be trying to keep the style of the first picture without having to mention it in the prompt

transparency is perceived as a mask if selected

those are the two things that struck me as a newbee

[mezotaken]

I see.

1. Infinite canvas is not infinite in openoutpaint
2. SDInfinity use interrogation to ensure the same style.

Transparency as a mask is purely inside-job from local inpainting implementation, probably worth a separate issue, this is not connected to openoutpaint.

1. How long ago it was not infinite? looks like they implemented it about 3 weeks ago:

https://github.com/zero01101/openOutpaint/pull/108

2. About keeping style: which model are you using?

[Frief84]

sd-v1-5-inpainting.ckpt ... I find stablediffusion-infinityis a bit more user friendly at least for me

[ClashSAN]

This is fulfilled by openoutpaint extension and blinkdl's (painthua ?)

[mikkorantalainen]

Am I missing something but isn't this simply better UI for inpainting? Basically the UI should display to be rendered tile area (512x512 for low VRAM people) which you can position partially overlapping the existing image and then write a prompt to tell what the area should contain. Then the missing part is automatically masked, the whole to-be-rendered tile area is used as input for img2img and ControlNet with canny/canny model + "pixel perfect" option and position the output when ready.