Outpainting 0.3 - thru img2img alt without a custom model, was gonna test it more but looks like i got scooped

[EtherealIntellect]

https://imgur.com/a/hMrJOe1

Step1: have an empty prompt generate an unconditional image, (promptless txt2img,they usually look like "stuff") . This up here was euler 125, keep the seed to copy to img2img, keep the resolution at what the final should be

Step2:put the original as small in the middle in a photo editor

Step3: Image to image alternative with no prompt for img2imgalt, just a single prompt for the reconstruction (i got: a painting of a mountain landscape with castles and hills with a sky background, by Jorge Jacinto - from clip). Deconstruct steps 125, reconstruct 125, denoising 1, reverse cfg shouldnt matter cuz empty prompt.

Step4: Combine the result back in a image editor, possibly edit the colours and blend it better than i did here- this was just a demo.

It has issues with cfg blowing out the colours, the middle image getting slightly changed in the reconstruction (no balloon and gate), negative prompts changing the image reconstruction too much etc. I was gonna test more - but testing this takes forever with how long the processing is, and might as well put it out there now that the seemingly better way is out.

Also, some images are bound to resolution - you can't outpaint the mona lisa without increasing resolution for example - it will just redraw a full mona lisa out of the noise painting over the middle small pic completely.

Imgur album too - they're a recreation of the other post https://imgur.com/a/hMrJOe1

https://www.reddit.com/r/StableDiffusion/comments/y8m2ve/there_is_a_new_model_that_brings_sd/

Also can probably be used for inpainting and background replacement too but i haven't tested it much

[Ehplodor]

The "stuff trick" is literally crazy good for outpainting imho !

[EtherealIntellect]

The "stuff trick" is literally crazy good for outpainting imho !

Yeah :) i spent way too long thinking about how to tell the AI - repaint this area, what would be "empty" since white was really sticking to things, and after a few weeks of ruminating on the computerphile video I remembered the uncond was literally the thing the ai was made to move away from. Idk what NAI uses - and havent tried the img2img alt noise slider too much, and I've seen some people use rgb noise, but ive had this come out more natural than rgb noise. The regular img2img uses "latent noise and latent nothing" - but i feel those partially stuck around in the final image too - only this got remade.

[Ehplodor]

Seeing the final result of your composite image gives me an idea: to achieve better integration, the colours should probably be corrected one way or another, applying colour correction to the final image using the copies of the original cropped pictures, or to the original cropped picture in a kind of reverse colour correction using could of the final image (before compositing)