[Basic Tutorial] Using Stable Diffusion XL (Patreon)

Hello, hello!

Note: ASK QUESTIONS IN THE COMMENTS IF YOU WANT, I CAN PLACE THE ANSWERS IN THE POST IF THEY ARE GOOD.

This is a short tutorial for you to start using Stable Diffusion XL.

I'll assume you are familiar with AI image generation.

If you never used Stable Diffusion before you can follow my tutorial for it. 

This is a very simple tutorial to clarify a few things about SDXL. I'll do more complex tutorials yet.

Stable Diffusion XL Model can be downloaded from Hugging Face and you can download

the VAE as well. If you want to skip downloading the model and only download the

VAE it's a good idea, because we are going to use other models.

You don't need a super strong machine to run it on your computer.

So even less than 8GB VRAM is possible, but you may run out of memory sometimes.

You can place the VAE model under the \stable-diffusion-webui\models\VAE folder.

Now you can visit a website like civitai.com and download your preferred model.

I'm going to use Blue Pencil XL as example. You can download any version you want, I'm using the one from the link.

You can place the model under the \stable-diffusion-webui\models\Stable-diffusion folder

Coming from Stable Diffusion 1.5 the big difference is that XL can generate higher resolution images and need less negative prompts.

It also generate better hands and faces.

The current best anime model for SDXL can only be used in NovelAI website.

The best idea would be to follow the model creators recommendations, but you can test things yourself.

You can't use Textual Inversions or LoRA that are trained for 1.5.

After starting your Automatic1111 webui, you have to change the checkpoint used and the VAE to SDXL as in the image below. (Check the top of the screen).

I will generate 1024x1536, but you can test with 1024x1024 or less. Or even more too.

YOU CAN DOWNLOAD THE IMAGE ON ATTACHED FILES AND PLACE ON PNG INFO TAB IF YOU WANT TO USE MY SETTINGS FOR IT.

There are plenty of quality tags and other models for you to use, but it's better to check the model page for more.

SDXL can perform really fast in image to image.

You don't need an upscaler to increase the size of your images, but you can use high resolution fix if you prefer.

In this case I've increased the size by 1.5 and used a denoising strength of 0.25.

This helps fixing some minor issues with the image.

Keep in mind that Stable Diffusion XL can generate really big images, I've generated an image of 3072x1024 before and it looked good.

There are Textual Inversions and LoRAs that works for SDXL as well. You can look for them on Civitai just remember using the filters and select only SDXL 1.0 models.