My Setup for Testing Checkpoints & Loras (A Guide)
This is an EXTREMELY in depth article and guide towards the tools, tips and tricks i've learned across the years. Some text IS SHARED from Guy90's article, due to the local generation option being easier to copy from his article rather than fudge it from my lack of knowledge. His article is here: https://civitai.com/articles/759. My settings and setup for renting GPU's and rolling my stuff is a little different than his. Read on if you're interested in finding out.
Before that join up the Ktiseos Nyx discord for more updates on our loras and content as well as AI/ML Discussions. https://discord.gg/HhBSvM9gBY
KoFi if you want more in depth AI & Art Commissions: https://ko-fi.com/duskfallcrew/
Merch Shop: https://duskfallcrew-shop.fourthwall.com/
Stable Diffusion
scratches head .... This one can't be as short as Guy90's considering there's no way I can lie to you and say I only use one tool.
Local Generation Users (Automatic1111)
To start if you haven't already, please have Automatic1111. If you need help downloading check out this guide from Wrench1815--> https://civitai.com/articles/1722/the-no-way-to-fuck-it-up-this-time-guide-to-installing-auto1111-or-forge-or-the-proompt-ninja
That's it, pretty much. IF you're running AMD, please note that might be covered in Wrench's guide or Guy90's original article i'm unsure. AMD cards do run Auto, Comfy and otherwise. Forge installs much like Auto, so please feel free to read up on that via above.
Local Generation Users (ComfyUI)
Hooboy, I don't use this a lot IF AT ALL ... so right now I don't have a local installation guide nor the replacement for if you have an AMD card. If i can find you one - I'll GLADLY link it here.
Renting a GPU and getting access to Comfy, Forge or Auto
If you're not running a decent tower with the latest gear, or like me are running a potato Imac:
You can use my affiliate link to VastAI here: https://cloud.vast.ai/?ref_id=70354
This allows me to earn back a small amount to use so I don't have to put as much INTO my account.
You'll note that you will have access to: https://cloud.vast.ai/templates/ - Which gives you a guide to what templates are available.
If you want direct access to the version i've got for a template it's here: https://cloud.vast.ai/?ref_id=70354&creator_id=70354&name=Auto%201111%20Webui%2013.0.0%20(Runpod)
Extensions & Suggested Add ons for Automatic & Forge:
Currently I run a list of the things I use, and it's easier if I just sort of dump the links, and explain roughly what each one of them does.
adetailer (This is the Detailer tool most of us use, it takes Yolo V8 models and will detail/redraw small amounts of a picture and clarify it) https://github.com/Bing-su/adetailer.git
sd-webui-supermerger (This merges loras, checkpoints etc) https://github.com/hako-mikan/sd-webui-supermerger.git
stable-diffusion-webui-model-toolkit (Auto pruning, and other goodies) https://github.com/arenasys/stable-diffusion-webui-model-toolkit.git
batchlinks-webui (Batchlink things like Huggingface and other downloadables) https://github.com/etherealxx/batchlinks-webui
SD Web UI Lobe Theme: (Make your AI setup pretty) https://github.com/canisminor1990/sd-webui-lobe-theme
Embedding Merge (Textual Inversion Merging): https://github.com/klimaleksus/stable-diffusion-webui-embedding-merge
Civitai Browser + (Browser for Civitai) https://github.com/BlafKing/sd-civitai-browser-plus Danbooru Prompt: https://github.com/EnsignMK/danbooru-prompt.git
Untitled Merger: (An advanced WIP merger by Silveroxides) https://github.com/silveroxides/sd-webui-untitledmerger.git
VAE
SD 1.5: You'll note SOMETIMES we use the Earth & Dusk Vae, and that it's not clear that its' as effective as other VAE. This is because it's a sidestep or merge vae that's like 3% other vaes and 97% Waifu VAE/KF L VAE. There's so many to list, and a lot of the time it depends on your preferences on this to this day I'm not going to argue what works best because I do not even recall.
SDXL, Pony, Illustrious, Animagine & Noob: I prefer as of recent to use the Luna XL Vae, as this works with IllustriousXL, I really love the vibrance it brings. Most Illustrious models and PonyXL models come with VAE's baked in. Mine have recently started being baked with LunaXL, and this is because of a color preference. Some Vaes are more pastel, grey or otherwise. PonyXL just takes the standard SDXL vae and should work even with most XL vaes.
Flux/SD3 etc: Don't have a clue if they use a built in one, or need one or what. Ask me later... muuuch later.
Upscaler
This is something that i'm not sure how much different my choices effect the output, because i'm weird and I use different ones. So i'd rather give you Guy 90's details which are more in par with our peer's usual dealings. I experiment alot, so it's best you get a handle of the basics before being like me.
Here is a website for more options: https://openmodeldb.info/
Embeddings
Ah, the thing that starts arguments on the kitchen table.
SD 1.5 embeddings DO NOT GO ON SDXL, ILLUSTRIOUS AND PONYXL.
First of all it's because the text encoders YES are dual on SDXL base models, but it's also that the two text encoders are TRAINED entirely different than SD 1.5. So even if you had dual embeddings to fit both encoders, your old SD 1.5 ones will not work properly.
Yes, because they're the same base architecture, they DO work "IN A WAY" - but most of the time they just crunch your image out and make it look weird.
Read more about this here: https://civitai.com/articles/10566/the-hidden-truth-about-textual-inversions-in-sdxl
If anything, I have my personal ones I use for all of them, but I'm not going to sit here and advertise ones that are my own, you can search for whatever.
One thing you need to note is: Embeddings work differently than loras, and don't usually change the style MUCH. Trained embeddings seem to do this a little more than Embedding merge type embeddings on SD 1.5. However, iv'e not seen much of a difference on XL based model sets.
If you'd like to see Embedding Merge in action and see how you can make your own watch my recent VOD here:
Merging & Lora Fusion
This is one that isn't always covered much, but as you note there's two plugins/extensions listed above. Untitled Merger doesn't have a lora function as of yet, but has some more advanced MBW settings that Supermerger does not. If you watch the above VOD that i've linked in the Embeddings section, you'll see Supermerger in action, and how to use it.
I recommend that when you're using LORAS to bake/fuse into your model that you vary the weights enough to not override the text encoders, or rather not terribly FORCE the change too much of the model.. This is because you can sometimes overcook a model by forcing too many high strength loras in.
This is because sometimes your best option sadly is and I will admit this myself: IF you're going for a strong style, you may want to consider learning to train the model instead of lora fusion. Lora fusion is great for the most part, but it's not efficient, and sometimes can burn and cook a model in the wrong way.
I've had luck most of the time, but sometimes my models suffer in the name of influencing the style.
Adetailer & Others
I haven't always been using Adetailer as much as I could, because on XL based models it takes longer and i'm impatient. However it still plays an absolute VITAL role. Sadly some of the best of these have been "BANNED" off the site since Anzch has been removed from the site after the creator program beta was removed from the site.
Their YoloV8 models can be seen on Huggingface here: https://huggingface.co/Anzhc/Anzhcs_YOLOs
You can use batchlinks to access them.
Online Generators (Not Local or Rented)
Small bits of options here and there besides just Civitai's generator!
Diffusers is evidently still running on Google Colab, as some ComfyUI users are dodging the AUP for this. Hell, some people are trying to do the inevitable and still run 30 accounts running Automatic1111. (The original reason why Auto1111 got banned on Colab in the first place.)
I don't link to Colabs for generation, because I don't want to get you in the risk of your account getting terminated. Lora training and other things seem to still be OK, and i'm not saying ALL free generations are "BANNED" - just I prefer not to mess with Google with the fact i'm all tied up with an android phone, storage etc...You CAN run free things via Amazon Sagemaker still, but that's a risk on you and I have never tested it. I've heard it's slow, and you get VERY LITTLE space to work with.
SeaArt (Yes, yes I get it , they "STEAL" but they're just ... automating a bot to sync your content, lmao stop abusing their staff XD)
TensorArt: I'm back, and yea i had some issues - Generator is mostly decent!
Moescape/Yodayo: https://moescape.ai/@duskfallcrew
ShakkerAi: Personally don't bother, not because their site isn't nice or their staff aren't nice, but the site is slow and the users will steal and claim your content as their own and not even credit you.
There are MANY other paid and free generation sites, and if you want me to test and rate yours after two years of being in this game: All I ask for is either a place to upload my content to your site OR if you're somewhere like DiffusionHub or otherwise that has rentables with space: I usually grab 120-150gb of space at a time, and if y'all got reasonable up/down speed and let me have a few hours of your time i'll be a happy camper!
Conclusion
Everyone is entitled to their own opinion and I AM ABUSING my own privileges. No, I'm just joking. Basically everything is variables, and everyone's got different things. I will be absolutely be updating this article, and editing it, maybe even cleaning up the lingo. I started writing this in the middle of playing Final Fantasy XIV.
