MultiDiffusion VS sd_lite

For composing this approach works pretty well

https://multidiffusion.github.io/

So I've continued to experiment with how many papers I can fit into a single pipe and have them play nicely together. The images below were created by combining the panorama code from omerbt/MultiDiffusion with the ideas from albarji/mixture-of-diffusers. Also turns out nateraw/stable-diffusion-videos can be seen as a special case of a panorama (in latent space rather than prompt space).

Essentially, I fine-tuned Stable Diffusion 2.1 base (the 512x512) model on the ESA Hubble Deep Space Images & Captions dataset I collected from public Hubble images & captions. After around 33,000 training steps, I saved the model and was really impressed by the results. But I really wanted to be able to generate wallpaper-level quality space images, so I stumbled upon MultiDiffusion: a new project for generating massive panorama images using stable diffusion models. I then used hubble-diffusion-2 along with MultiDiffusion to generate each one of these amazing 2560x1536 images. Each image took a little over an hour to generate on a Google Colab T4 GPU. I used the following prompts for each of these images:

If you need generated and not upscaled 4k for some reason, try something like https://github.com/omerbt/MultiDiffusion

Project: https://multidiffusion.github.io/ Paper: https://arxiv.org/abs/2302.08113 GitHub: https://github.com/omerbt/MultiDiffusion

Quick Read: https://www.marktechpost.com/2023/02/24/meet-multidiffusion-a-unified-ai-framework-that-enables-versatile-and-controllable-image-generation-using-a-pre-trained-text-to-image-diffusion-model/ Paper: https://arxiv.org/abs/2302.08113 Github: https://github.com/omerbt/MultiDiffusion Project: https://multidiffusion.github.io/

So I haven't made many images with Stable Diffusion despite using it heavily. The reason is I've been messing with the internals of the diffusion pipe, to interfere with the diffusion process in different ways. Todays fun result is based on omerbt/MultiDiffusion for making panoramas.

If we combine Mixture of Diffusers + MultiDiffusion+ Composer+ cross-domain-compositing and probably some more I'm not thinking of.

The full details of how to do this are available on Github: latent jitter · thekitchenscientist/sd_lite but I will explain the idea briefly here. I have read this could be done with perlin or simplex noise but the code was too complex for my taste. This gets the job done with only minor modifications to the standard pipe.

I did successfully swap the effiel tower for the burj Khalifa but that required additional steps https://github.com/thekitchenscientist/sd_lite/wiki/latent-jitter

I use it in my pipe as a general image beautifier. https://github.com/thekitchenscientist/sd_lite/wiki/safe-latent-diffusion

The pipe is available at sd_lite/pipeline_stable_diffusion_multi.py (github.com) it combines:

This community has published many XY plots of CFG versus steps. https://github.com/thekitchenscientist/sd_lite/wiki/recommended The consistent theme is low CFG, lower steps; high CFG, more steps. UniPC can reach convergence in as few as 8 steps, so I increased by 1/3 to account for more complex prompts needing longer

the pipeline is available from github.com and is called in the usual way. The Technique requires the DDIM scheduler.

https://github.com/thekitchenscientist/sd_lite is where the code is. Version 1 of the multi-pipe is limited to images 512 high or wide but any number on the other dimension

Not to be deterred I hacked together some code to blend it all back together after the VAE but before the final colour balance. The pipe code is available on github. thekitchenscientist/sd_lite