db0

One-Click Worker Install, Multiple Models and img2img, oh my!

Added 2022-10-19 10:57:03 +0000 UTC

I have started a closer collaboration with hlky and the (soon to be renamed) sd-webui group, on [a unified backend for SD generations](https://github.com/sd-webui) and the first thing I did was port the horde bridge into it.

The second thing I did was ask henky from the KoboldAI crew to help me setup a one-click install of the thing using micromamba. This means that you can new head to this the [nataili repo](https://github.com/sd-webui/nataili), download the zip or git clone it, and run `horde-bridge.cmd` or `horde-bridge.sh` and you're good to go! No more installing 3-4 different tools and then the models manually, and hoping nothing broke in the process!

The great thing about the nataili backend is that it is way more modular than working on the UI as a backend, so I can be way more flexible with the components I'm using in the bridge. This fact enabled me to very quickly add img2img support to the worker and therefore the Horde!

Unfortunately due to the ethical problems posed by anonymous accounts sending potentially illegal images to random workers, this feature is only available to trusted user of the Horde, which is anyone who has onboarded their worker to the horde and has helped for a significant amount of time, as well as individually vetted tool integrators. I hope to unleash the img2img to the general public soon.

And last but not least, the new backend allows me to easily support different models per worker, as well as multiple models per worker! This allows people with fancier GPUs to provide even more flexibility to the horde by covering more models at the same time, and people with specific niche interests and a decent GPU can support their individual communities by hosting their in-demand models for the rest to use!

The ability to host (multiple) models other than SD is, from what I understand no other online AI image generator other than the horde allows. And certainly no free ones I've seen!

These two new features combined, allow for some really cool options. Not only things like random models per generation, but also changing the same image into multiple styles very easily! Not even the most powerful GPUs can adjust an existing image to 10 different models so fast.

Here's an example of me converting my avatar to different models, using img2img. This whole process took me like 1 minute (and this is with 5 images per run, of which I chose only 1 to show here)

Prompt: grinning wild west outlaw walking towards the camera, gunbelt, detailed face, explosion in background