Improve Outpainting

[cerulean inlet]

I've looked into outpainting and was wondering, why it right now replaces parts. For me that's more inpainting than outpainting. 🙂
Is there a reason for it? I'm happy to open a PR to extend the functionality to create bigger images. Means instead of moving it, we extend the image and it could get bigger and bigger.

Not sure what the limitations would be. I guess we should still just put a piece of it into the generator if we do it multiple times.
Example:

• 512x512 image
• outpainting top 50
• ends in 562x512 image
• continue using the new image with 562x512
• outpainting top 50
• using 512x512 of the 562x512 for the generator and add it to the 562x512
• ends up in 612x512

WDYT?

[cerulean inlet]

closing as @neon sapphire wanted to look into the topic anyway. @neon sapphire let me know if you need help.

[neon sapphire]

I could always use help and would welcome a PR! Please be advised that the current algorithm works by generating a mirror image of the original such that img2img has a starting place that looks similar in terms of colors. See https://github.com/invoke-ai/InvokeAI/pull/489. The part that I'm planning to look into is whether initializing the crop area with latent noise rather than a mirror image will produce better results. I'm not doing anything about changing the image size.

[cerulean inlet]

@neon sapphire thanks for the PR link. helps a lot to understand the implementation right now. funny Any-Winter-4079 also played with merging the images.
That's what I'm wrapping my head around right now.

There are so many possibilities with out- / inpainting:

• specify with what it will be in-/outpainted (if you add the noise like you described, I could better guide with a prompt what I would expect)
• enlarge images through outpainting and enhance feature / details

I have something in-mind like this: https://www.reddit.com/r/StableDiffusion/comments/xelydv/stablediffusioninfinity_outpainting_with/

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

[cerulean inlet]

@neon sapphire and this one 🙂 https://www.reddit.com/r/MachineLearning/comments/xnbv8e/p_enhancing_local_detail_and_cohesion_by/?utm_source=share&utm_medium=ios_app&utm_name=iossmf