Reinstalling Stable Diffusion (Patreon)

"This is seiso and important information, Author-kun!"

If you read my Work Notes, you know that I borked my SD installation, so I spent the first half of the week reinstalling it. It may seem scary at first, but reinstallation is really easy!

Don't delete the old folder... Yet! - If space allows, you shouldn't delete your old files completely. You can reuse the checkpoints and LORAs that are contained there, not to mention a treasure trove of generated images. It would be useful to keep even the discarded Work files if only to refer back to the data later. Checkpoint Models, LORAs and generated images! Keep these things!

Installation is as simple as a git clone - There isn't much use for an all-in-one installer since the only thing you have to do is just create a new folder, check there's enough space (maybe 15gb of free disk space?), and open up a command prompt to do a 'git clone'. I say this because all the dependencies have already been installed, your mileage may vary if installing for the first time.

Interesting Changes

Refiners for SDXL - I don't really use SDXL, but apparently this is the 'high-res fix' for SDXL models. I'll have to look into that after I finish working on the Lewdicron this week.

Render Preview is built-in and enabled  - You used to have to manually turn this on, live render previews are not the default. Gone are the days where you are caught wondering if your image was going to come out of high-res fix half-baked!

LORAs are moved to a separate tab - I heard of this news a while ago but didn't update. LORAs are much easier to use now, you just open up the tab (moved from below the 'generate' button) and it detects Lycoris and LORAs alike. I still very much enjoy the old AdditionalNetworks interface

Default Denoise Strength is 0.5 now - One of the weird things about high-res fix is that it used to be set at a very high denoise by default. Now SD sets it to 0.5 as a default, which I've found from my testing that it keeps the base image most of the time, as with anything <0.5 denoise strength. I usually turn it up to 0.6 for more details.

PNG Info - I didn't have this in my old installation; its a new tab where you can just check for the details of any Work you create. Just browse to the image and it will read the prompts it was created with! (if metadata exists)

Essential Fixes

Change Clipskip to 2 - Since I mostly work with anime, I change clipskip to 2 to get a more 'anime' feel to the Work.

Install Extension: Image Browser - Sometimes I just want to browse through Work that I have done for the day, and using PNG Info to look up prompts one by one doesn't quite cut it. Image Browser allows you to look at all the Work for a day at a glance, and pops up generation data when you

Install Extension: ControlNet - You can't get away from ControlNet if you are looking to do Work. Even if you are not into skeletons like Author-kun is and into poses like I am, it doesn't hurt to mess around with just the openpose module at least. It truly is a magical feeling when the characters pose EXACTLY the way you want them to! (you have to download the models separately, I just moved them from the old installation)

Install Extension: sd-webui-tagcomplete - To be honest, I was doing the Utena pics without this, and it was really difficult. tagcomplete helps you standardize your tags like a spellchecker, but for danbooru tags. So 'arms_bound' will be corrected to 'bound_arms'. It allows me to see how popular/consistent the tag is. I consider it essential for making good images.

Last Word

I am currently keeping both installations on the computer and slowly moving over LORAs as I need them. I'll delete the old folder when I don't need anything from it anymore. Just wanted to put it out there that SD installation/re-installation is not as scary as it seems!

GMVio

240111