Checkpoint: Base Model

The first time you see Checkpoint in ComfyUI, you might wonder: is this the model itself, or some kind of plugin?

One sentence to remember: A checkpoint is the “main painter” — the foundation.
Everything you add later — LoRA, ControlNet, VAE — builds on top of this foundation, not a replacement for it.

Understanding Checkpoints in Plain Terms

Think of an image generation pipeline as a team:

• Checkpoint: The main painter (sets the base style, aesthetic, and capability ceiling)
• LoRA: A style plugin (temporarily adds skills to the main painter)
• ControlNet: A pose director (constrains composition / pose / edges)
• VAE: A translator (converts latent space into a visible image)

So with the same prompt, swapping the checkpoint is like having a completely different person do the painting.

Where to Put the Files

Default location:

ComfyUI/models/checkpoints

Once placed there, you can select it in nodes like CheckpointLoaderSimple.
If you can’t see the file in the dropdown, check three things:

1. Wrong path (most common).
2. Incorrect file extension (should be .safetensors).
3. ComfyUI may need a restart or a model list refresh.

Practical Tips for Choosing a Checkpoint

Don’t start by hoarding dozens of models. Get 1–2 solid base models working first, then expand:

• Realistic output: Choose a realism-oriented model.
• Anime / illustration: Choose an anime-oriented model.
• Product / e-commerce: Prioritize models with stable structure and clear detail.

Core principle: pick a model pointing in the right direction first, then tune parameters.
If the base model is pointing the wrong direction, no amount of prompt tweaking will fix it.

Checkpoints vs. LoRA (the Most Common Confusion for Beginners)

Many people mistakenly think LoRA can fully replace a checkpoint. It can’t.

• Checkpoint is the foundation
• LoRA is an add-on

If the foundation is realistic and you attach a heavily anime LoRA, you can pull the style, but there’s usually a tension.
Common advice: LoRA and the base model should share the same architecture and style family.

Checkpoints vs. VAE

Some checkpoints include a built-in VAE; some workflows load a separate VAE.
If you see “washed-out, muddy, dull, or weirdly-colored images,” first suspect:

• Whether the VAE matches the model
• Whether the workflow is loading the VAE twice or incorrectly

This is a very common issue — not user error.

Common Pitfalls

1. Only looking at the preview image, not the compatibility notes: You download it, spend time setting up, then can’t run it.
2. Mixing SD1.5 and SDXL: Architecture mismatch causes immediate errors or garbage output.
3. Files in the wrong folder: The file exists on disk but outside ComfyUI’s scan path.
4. Stacking too many LoRAs at once: Styles fight each other — looks like “everything and nothing.”

A Minimal Exercise You Can Do Right Now

1. Pick your most-used checkpoint.
2. Fix the prompt and seed, run 3 images.
3. Swap only the checkpoint, run 3 more images.
4. Compare — you’ll quickly build intuition for “what this model is good at.”

This exercise beats reading ten reviews.

Further Reading

• Models & Files
• Using LoRA
• Comfy docs index