What Are Stable Diffusion Checkpoints? + How To Use Them
- Cole B.
Today, I am going to cover everything you will need to know about stable diffusion checkpoints.
This guide will teach you how to get started using this integral feature of stable diffusion.
Let’s dive right in:
What Are Stable Diffusion Checkpoints?
Stable diffusion checkpoints are models trained with huge amounts of images with text descriptions, similar to prompts.
This allows the model to know what keywords would be associated with each image. When you generate images using the same keywords, the model will generate an image similar to the image dataset it was trained on with the same keywords.
The image dataset that was used to train the checkpoint dictates the style of the images it produces and which keywords are understood. Checkpoints usually have a comphrensive amount of keywords, so you shouldn’t have to worry about researching which ones to use.
Smaller models, like LoRAs, tend to be much more specific with their keywords because they are trained with much less data.
How To Use Checkpoints
Now, I am going to address generally how to use checkpoints, although each stable diffusion software will be different:
1. Download the file
First, I recommend visiting either Civitai or you can use Huggingface to find different models that you would want, then download them.
2. Upload it to the right folder
This is the part that is going to differ for you. Depending on which software or program you are using, you will need to move the file to a different folder.
I am going to show an example for the folder of Web ui (automatic1111):
stable-diffusion-webui/models/Stable-diffusion/
Once the model is moved to the correct folder, you can move on to the next step.
3. Reload stable diffusion
Most software will require you to reload the UI for the checkpoint to appear as a selection, so I would recommend closing and opening the program you are using.
4. Start using your new model
Finally, you should be able to use the model that you have downloaded and start making AI art.
Different File Formats
When downloading models, you are sometimes given a few different options of file types:
- Safetensor
- Pickletensor
- Diffusors
I am going to walk you through the differences and which one you should download, because some file formats might give you problems in the future.
Safetensor (.safetensor or .safe)
Safetensor is the preferred file type you should download if there is an option. They have a few advantages:
- The serialization format of safe tensor files ensures the integrity of data and decreases the risk of downloading malicious code.
- Safetensor files have a smaller file size than pickletensor files. Just a small upgrade but still worth it if you are downloading a lot of checkpoints because they are the largest stable diffusion model.
Pickletensor (.ckpt)
Pickle tensor files are not recommended because they have the potential to download malicious code to your computer.
Only if you have to, you can download pickletensor files, and you should scan them with a anti-virus program.
Diffusors
Diffusors are just a bunch of files that are not packed together for the checkpoint.
The other file types come packed together in a zip file, but this one does not. Nothing wrong with using diffusors, but safetensors are best for most people.
Best Checkpoints For Each Image Style
Anime Ckpts
All of these models will produce great-looking anime outputs. To compare and see which one you may want to use, I am going to use the exact same prompts and parameters and show the results in a 2 by 2 image grid.
- Prompt: anime, girl sitting on a hillside, daytime, sunny, best quality
- Steps: 40
- Sampler: DPM++ 2M
- CFG Scale: 7
- Size: 512×512
Anything XL
Anything v5 or XL is one of the most popular anime image models.
Rev Animated
Rev animated is more multipurpose than just doing anime images. This model can do fantasy, semi-realistic, and landscape images as well.
Mistoon
Mistoon produces more saturated and bright anime images compared to the others.
AingDiffusion
Lastly, AingDiffusion makes images in a slightly less modern style by default but can be experimented with. Can also look similar to American animated cartoons that mimic the anime style.
Realistic Ckpts
Next I am going to show you models that are good for producing realistic images.
The settings I used are below:
- Prompt: adult man, portrait, realistic, photorealistic, photo, dark hair
- Steps: 40
- Sampler: DPM++ 2M
- CFG Scale: 5
- Size: 512×512
JuggernautXL
JuggernautXL has to be one of the best all-around stable diffusion XL models. It can do most different styles, but it does realistic the best.
Realistic Vision
Realistic vision is a solid choice for human subjects, but it does look slightly more fake than the next two.
epiCRealism
epiCRealism does a great job of making realistic-looking images.
CyberRealisim
CyberRealism has a similar realism to epiCRealism, with maybe a more moody look.
AbsoluteReality
AbsoluteReality has real-looking facial features, but the texture might seem a little fake.
Where to Download Checkpoints
Hugging Face
Although hugging face isn’t specifically for stable diffusion, it has lots of information over AI-related topics in general. It can be a great place to download and find new models or discussions on AI news.
Civitai
On the other hand, Civitai is meant for stable diffusion and AI image generation. You can find new models and images from other artists very easily. Probably the best website for downloading what you need for stable diffusion.
I’d Like To Hear From You
Now that I have gone over most things regarding stable diffusion checkpoints.
Is there anything I missed?
Let me know by leaving a comment below.