PROMPT GUIDE

How to Choose a Stable Diffusion Model (Checkpoint)

In Stable Diffusion, the look of your images is decided more by which model (checkpoint) you use than by the prompt. There are countless models available, and it is easy to get lost. Rather than naming specific models, this guide explains how to think about choosing one and what to check.

Last updated: 2026-10-22 approx. 3 min read

日本語版

The short answer

Choose a Stable Diffusion model by three things: what kind of images it is good at, which generation it belongs to, and the terms under which you may use it. Check the recommended size, quality tags and license on its download page, then test a few images with the same prompt before making it your main model. Prefer the safetensors file format.

What is a Stable Diffusion model (checkpoint)?

It is the core file for making images, and it decides most of the art style and what the model is good at.

A checkpoint is a large file that bundles what was learned from a huge number of images. With the same prompt, a model good at anime gives you anime art and a model good at photos gives you photo-like images. A LoRA only adds a few features on top of that model, so choosing the base model comes first.

Type	Best at	What to look for
Anime / illustration	Anime coloring, characters	Often written with Danbooru tags. Check the recommended quality tags
Photo / realistic	Photo-like people and scenery	Easy to prompt with sentences. Mind the rules on depicting people
2.5D / semi-realistic	Between illustration and photo	Skin and light textures come out well
Background / scenery	Landscapes, buildings	Better at backgrounds than people
Merged models	Features of several models	Also check the licenses of the source models

Anime-style girl in a school uniform, an image of an anime-oriented model — アニメ・イラスト系のイメージ

Anime girl in a flower field drawn with soft painterly colors — アニメ・イラスト系のイメージ

Illustration usage rules

What should you check when choosing a model?

Start with four things on the download page: generation, recommended settings, license and file format.

What to check	Where to look	Why it matters
Generation (Base Model)	SD1.5, SDXL, Illustrious, Pony, etc.	Decides which LoRAs and VAEs are compatible
Recommended size and settings	The Recommended section of the description	Images break more easily if they don't match
Recommended prompts	Example quality tags and negatives	Gets the look the model intends
License and terms	The License field and description	Commercial use and other conditions differ
File format	.safetensors or .ckpt	safetensors loads more safely
VAE included or not	Notes like VAE baked	A common cause of dull colors

Rinka's tips

.ckpt files can contain program code, so avoid ones of unknown origin and choose the safetensors format
Some models are also released in lighter versions such as fp16. The look is nearly the same and the file is smaller
Licenses can change with updates. Check again before use and before any commercial use

How do you test a model?

Steps to try a new model

With your current model, pick one reference prompt and seed and save them
Put the new model in the models folder (Stable-diffusion or checkpoints) and select it in the interface
Match the recommended size, quality tags and negative prompt from the download page
Make a few images with the same prompt and seed and compare them with your previous model
Check that the images you usually make, including expressions, hands and backgrounds, don't break before adopting it

Reference prompt example: 1girl, solo, long hair, smile, school uniform, classroom, window, sunlight, upper body, masterpiece, best quality

Comparison note example: model name / size 832x1216 / seed 12345 / quality tags: as recommended / notes: hands stable, background bright

For how to fix the seed and compare, see What Is a Seed? Fixed, Random, and Reproducing the Same Image →

To compare several models in one grid image, see How to Use the XYZ Plot (Compare Settings Side by Side) →

How should you choose an anime model?

Pick a generation you can prompt with Danbooru tags, and start with models that have many examples in a style you like.

Many anime models in use today are based on SDXL, such as the Illustrious and Pony families. Recommended quality tags and prompt styles differ by model, so if sample prompts are published, start from those and adapt them to your own images to get used to the model quickly.

For the prompt order in Illustrious-based models, see Illustrious and NoobAI Prompt Order →

What are common mistakes and how do you fix them?

Problem	Cause	Fix
The model doesn't appear in the list	Wrong folder or list not refreshed	Put it in the right folder and refresh the list
Images are black or colors are wrong	Mismatched VAE or half-precision issue	Use a VAE that suits the model and review settings
Results don't look like the samples	Size or tags differ from the recommendation	Match the recommended settings before comparing
A LoRA doesn't work	LoRA and model generations differ	Pick a LoRA of the same generation
You can't decide which model is best	Comparisons were inconsistent	Compare with the same prompt and seed

For SDXL vs SD1.5, see SDXL vs SD1.5 (Image Quality, Image Size and LoRA Compatibility) →

For what a VAE does and how to choose one, see What Is a VAE? Why Colors Look Dull, and How to Choose and Set a VAE →

FAQ

Frequently Asked Questions

What is the difference between a checkpoint and a LoRA?

A checkpoint is the large base file for images, and a LoRA is a small add-on that adds a few features to it. Use LoRAs that match the generation of your model.

Should I choose safetensors or ckpt?

We recommend safetensors. A ckpt file can contain program code, so avoid ones of unknown origin.

Where can I check whether a model can be used commercially?

Check the license field and description on the download page. For merged models, the terms of the source models also matter, so check those too.

I installed a model but it doesn't show up.

Check that it is in the right folder and that you refreshed the list in the interface. Folder names differ by interface.