What is a checkpoint in AI image generation?

A saved set of model weights — the file that actually generates images. Different checkpoints of the same architecture produce different output from the same prompt.

What is the difference between a checkpoint and a LoRA?

A checkpoint is the entire model, several gigabytes, and you load one at a time. A LoRA is a small add-on applied on top of a checkpoint, and several can be stacked.

Why do I get a different image from the same seed?

Because something else changed — most often the checkpoint. A seed only reproduces an image when the model, prompt and settings are also identical.

How big is a checkpoint file?

Typically several gigabytes, which is why storage becomes a real consideration once you keep more than a few.

Can I use any checkpoint with any interface?

Only if the architecture matches. A checkpoint for one model family will not load in tooling expecting another.

Why does the word "checkpoint" mean a model file?

Because weights are written to disk periodically during training so progress survives a crash, and the file people download to run is one of those saves.

Are community checkpoints safe to use commercially?

Not automatically. The base model licence still applies, distributed checkpoints carry their own terms, and the training material behind a fine-tune is often undocumented.

AI image glossary

What is a checkpoint?

A saved set of model weights. Different checkpoints of the same architecture produce visibly different output from identical prompts.

A checkpoint is a saved snapshot of a model’s weights: the multi-gigabyte file that does the generating. Two checkpoints built on the same architecture will produce visibly different images from an identical prompt and seed, because the weights encode everything the model learned and each has been trained or fine-tuned on different material.

A row of graduated brass weights, each unstamped, on a warm paper surface

The name comes from training, not from images

During training, weights are written to disk periodically so that progress survives a crash and so that different stages can be compared. Those saved files are checkpoints in the ordinary engineering sense. The word escaped into general use because the artefact people download and run is one of those saves, and the label came along with it.

What makes two checkpoints differ

Almost always fine-tuning. Someone takes a general base model and continues training it on a narrower body of images — photography, anime, architectural renders, a particular illustrative register — and the resulting weights lean toward that material. The architecture is unchanged, so the same tooling runs it, but the output has a different default character and different competencies.

Checkpoint versus LoRA

A checkpoint is the whole model and runs to several gigabytes. A LoRA is a small adjustment applied on top of one, typically a few hundred megabytes at most. You load exactly one checkpoint at a time; you can stack several LoRAs on it. If you want a model that is fundamentally good at a genre, that is a checkpoint decision. If you want to add one style, character or object, that is a LoRA.

Why the same prompt and seed give different results

A seed fixes the starting noise, and identical noise through identical weights gives an identical image. Change the checkpoint and the weights are different, so the same noise is denoised along a different path. Reproducibility therefore requires recording the checkpoint alongside the prompt and the seed — a seed on its own reproduces nothing.

The practical caution

Checkpoint files are large, licences vary considerably, and provenance is often poorly documented in community distributions. A checkpoint fine-tuned on material its distributor had no right to use does not become safe by being convenient, and the licence on the base model continues to apply to anything derived from it.

Checkpoint at a glance

Is: A saved set of model weights
Size: Several gigabytes
Loaded at once: Exactly one
Differs by: What it was fine-tuned on
Needed to reproduce: Yes, with prompt and seed
Cross-architecture: Not interchangeable

Related terms

What a checkpoint sits next to

What is a LoRA?The small add-on on top.

reproducing an imageThe other half of reproducibility.

What is Stable Diffusion?Where checkpoints proliferated.

the model directorySeven models, described.

What is a sampler?How the steps are taken.

the glossaryEvery term, defined.

Put it to use

Where this appears

AI image generatorThe main generator — text to finished image.Open Prompt Gallery120 images with the exact prompts that made them.Open transform an existing imageStart from a picture you already have.Open edit part of a photoChange part of a photo, keep the rest.Open enlarge an imageEnlarge to 4K without the smearing.Open enhance a photoExposure, noise and old-photo restoration.Open remove something from a photoMasked removal that rebuilds the background.Open blur a photo backgroundDepth-aware, not a uniform filter.Open Colorize PhotoPlausible colour for black and white.Open generate a logoMarks that still read at 16 pixels.Open design a tattooDesigns that survive being tattooed.Open generate illustrationIllustration and painting, not photographs.Open generate a headshotStudio-grade portraits without a studio.Open the photo specificationThe exact specification, and how to meet it.Open interior visualisationRestyle a room you can photograph.Open AI Image GlossarySeed, denoise, inpainting, explained.Open

Checkpoint — questions people ask

What is a checkpoint in AI image generation?: A saved set of model weights — the file that actually generates images. Different checkpoints of the same architecture produce different output from the same prompt.
What is the difference between a checkpoint and a LoRA?: A checkpoint is the entire model, several gigabytes, and you load one at a time. A LoRA is a small add-on applied on top of a checkpoint, and several can be stacked.
Why do I get a different image from the same seed?: Because something else changed — most often the checkpoint. A seed only reproduces an image when the model, prompt and settings are also identical.
How big is a checkpoint file?: Typically several gigabytes, which is why storage becomes a real consideration once you keep more than a few.
Can I use any checkpoint with any interface?: Only if the architecture matches. A checkpoint for one model family will not load in tooling expecting another.
Why does the word "checkpoint" mean a model file?: Because weights are written to disk periodically during training so progress survives a crash, and the file people download to run is one of those saves.
Are community checkpoints safe to use commercially?: Not automatically. The base model licence still applies, distributed checkpoints carry their own terms, and the training material behind a fine-tune is often undocumented.

Related AI image terms

Start creating today

Generate your first image with the best AI image generator.

Turn a sentence into a finished, photorealistic image in seconds — then refine every detail. No setup, no Discord, no GPU.

Generate without downloading a model Get Started

Join 4,200+ creators using LaFoto