[Guide] Introducing SDXL-Lightning (Patreon)

What is SDXL-Lightning?

SDXL-Lightning is a lightning-fast text-to-image generation model. It can generate high-quality 1024px images in a few steps.

This definition is taken from here: SDXL-Lightning

The page from the link has more references and it has LoRA models that can achieve this effect.

Why use this model?

This can be an excellent tester, or it can also help slow pcs to make much faster images. The quality is not so different, but there is a noticeable difference in some small details so I recommend that you test and judge on your own.

I never used any AI image model where should I start?

This is not for people who are starting, you can visit my first basic tutorial and check more info in this collection.

If you don't have the GPU for that, but are interested in generating something you can try online services. My first recommendation is to use Midjourney, is a different AI, but simpler to get good results. You will have to use discord for it, but you can place the bot on your own server. It's a paid service, but you can try for free.

There is also the chat gpt dall-e 3, you can try that using the Microsoft Copilot.

Anyway, let's get going with this.

The LoRA:

You can use the LoRA and basically reduce the steps, but usually you should reduce the CFG if you feel like it's producing too much saturated images.

Ideally you should also use an updated version of Automatic1111 webui or something similar that have the option for "Schedule Type".

This will be useful because actually I want to give a hint on how to use a pre-trained Pony Diffusion model that has the LoRA merged to it

AutismMix SDXL:

This is an example of a model, but many people can merge their checkpoints with the model. You can do it yourself with your favorite checkpoint, but I'll not cover it now because I didn't test.

Here is the model link: AutismMix SDXL Lightning 4step

You can download the 4 step or 8 step version.

Example:

For this example I'm going to use this LoRA of Rin (Blue Archive), you can use whatever you want, I'm going to use the same prompts as one of the lora examples.

You can download the image from the gallery and use it on PNG Info tab from your Automatic1111 webui if you want to test it.

The difference is that I'm going to use less steps and CFG scale. Also the other important part which is told in the instructions of AutismMix SDXL 4-8 steps. Use schedule type SGM Uniform.

The model is called 4 steps, but you can use a bit more. I'm using 6.

My settings can be used from the example image as I told you, but basically I'm using:

Sampling Method: DPM++ 2M.

Schedule Type: SGM Uniform.

Sampling Steps: 6.

Width: 1024, Height: 1024.

CFG Scale: 2

Conclusion:

This is around 6-10 times faster, sometimes even more depending on what you are doing. It can be even more, because if you only depend on steps to be faster it will do the trick.

This is not a "how to make an image in style of..." so that's why the girl is not in any specific style that I could do. You can use a style lora in conjunction with this if you want.