In AI, a checkpoint is a saved snapshot of a model’s learned numbers at one point in its training, so the work can be resumed or used later. In AI image tools, the word also means the downloadable model file itself: the file you load to make pictures.
Think of a save point in a video game. Training a model takes weeks, so the makers save its progress along the way. If something crashes, they reload the last save instead of starting over. When a model is finished and shared, what you download is simply one of those saves.
In a sentence: “I swapped the checkpoint for one trained on anime art, and the same prompt came out in a completely different style.”
Two senses, one idea
The plain point: both senses mean a file holding a model’s learned numbers; they differ in who uses it and why.
| Sense | What it is | Who says it |
|---|---|---|
| Training checkpoint | Is: a save made during training, to resume or compare later. | Heard from: the people building models. |
| Model checkpoint (image tools) | Is: the finished or fine-tuned model file you load into an image app. | Heard from: people making pictures with open models. |
The training sense: a save point
The plain point: a checkpoint stores the model’s numbers, not the program that runs them.
A model’s knowledge lives in its parameters, the numbers it adjusts during training. Google’s TensorFlow guide says checkpoints “capture the exact value of all parameters” a model uses, and that they hold no description of the computation, so they need the matching code to be useful.1
A checkpoint meant for resuming training saves a little more. PyTorch’s tutorial says to keep the optimiser’s state too, plus extras such as the epoch you stopped on and the latest training loss.2
The image-tool sense: the model file
The plain point: in open image tools, “checkpoint” just means “the model”.
The two senses meet in plain sight. The original Stable Diffusion release listed its models as “checkpoints”, each one resumed from the save before it: sd-v1-2.ckpt, for example, was “Resumed from sd-v1-1.ckpt”.3 The name stuck. Hugging Face’s image guides still begin with “Load a checkpoint”, meaning load the model.4 ComfyUI, a node-based image app (see ComfyUI explained), tells users to put “your SD checkpoints (the huge ckpt/safetensors files)” in a folder called checkpoints.5
Many checkpoints are copies of a base model fine-tuned on one style or subject. Hugging Face’s own guide swaps in one made for a comic-book art style, noting a trigger word is needed “to use this checkpoint”.4 Swapping one changes the whole look of your pictures, unlike a LoRA, which is a small add-on loaded on top. Hugging Face notes that some add-on files are also called checkpoints, such as IP-Adapter ones of about 100MB, “because they store adapter weights, not a full model.”6
If an image tool asks you to pick a checkpoint, it is asking which model to use. If a researcher mentions a checkpoint, they usually mean a save from partway through training.
Is a checkpoint file safe to download?
The plain point: it depends on the file format and who made it.
Many older model files use a Python format called pickle. Hugging Face warns that pickle is the default for PyTorch model weights, including files such as pytorch_model.bin, and that “dangerous arbitrary code execution attacks” can happen when you load such a file: opening it can run hidden code.7 The newer .safetensors format was made to store the numbers “safely (as opposed to pickle)”.8 Prefer .safetensors, and download from makers you trust. ComfyUI explained covers the wider download risks.
Questions people ask
Is a checkpoint the same as a model?
In everyday use, yes: the checkpoint is the model’s learned numbers saved to a file. Strictly, TensorFlow’s guide points out that a checkpoint holds the values but not the code that uses them, so an app supplies the rest.1
What is the difference between a checkpoint and a LoRA?
A checkpoint is a whole model; a LoRA is a small add-on that changes it. Hugging Face says LoRAs work by training only a few new weights instead of the entire model, which makes them faster to train and easier to store.9
Should I pick .ckpt or .safetensors?
Pick .safetensors when both are offered: it was built as the safe alternative to pickle, the older format whose files can run code when opened. Hugging Face’s own software loads either kind of file.10
Learn more on this site
- What is inside: parameters (weights) meaning.
- Which files you may download: open weights meaning.
- Add-ons on top: LoRA meaning and fine-tuning meaning.
- Running models yourself: run AI free on your own computer.
- The words: the AI glossary.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. In AI image tools, what does ‘checkpoint’ usually mean?
- The model file you load to make pictures
- A button that saves your finished picture
- A safety filter that blocks some prompts
Show the answer
Answer: The model file you load to make pictures. In open image tools, ‘checkpoint’ just means the model: the file you load to make pictures.
2. Why do model makers save checkpoints during training?
- To make the model smaller for phones
- So they can reload the last save if something crashes
- To share each picture the model makes
Show the answer
Answer: So they can reload the last save if something crashes. Training takes weeks, so makers save progress along the way, like a save point in a video game, and reload it if something crashes.
3. Which file format does the page suggest picking when both are offered?
- .ckpt, because it is the older format
- Whichever file is larger
- .safetensors, because it was built as the safe alternative to pickle
Show the answer
Answer: .safetensors, because it was built as the safe alternative to pickle. Older pickle-based files can run hidden code when opened; .safetensors was made to store the numbers safely.
Related terms: AI winter · Alignment · Attention · Clustering · Computer vision · Constitutional AI
- 1TensorFlow (Google), Training checkpoints, last updated 23 Mar 2024, read at source 12 Oct 2026“Checkpoints capture the exact value of all parameters (tf.Variable objects) used by a model.” … “Checkpoints do not contain any description of the computation defined by the model and thus are typically only useful when source code that will use the saved parameter values is available.” ↑a“Checkpoints do not contain any description of the computation defined by the model”. ↑b
- 2PyTorch, Saving and Loading Models, last updated 26 Jun 2025, read at source 12 Oct 2026“It is important to also save the optimizer’s state_dict, as this contains buffers and parameters that are updated as the model trains.” … “the epoch you left off on, the latest recorded training loss”. ↑
- 3CompVis, stable-diffusion, GitHub README, undated, read at source 12 Oct 2026“We currently provide the following checkpoints:” … “sd-v1-2.ckpt: Resumed from sd-v1-1.ckpt.” ↑
- 4Hugging Face, Image-to-image, Diffusers documentation, undated, read at source 12 Oct 2026“Load a checkpoint into the AutoPipelineForImage2Image class”. ↑a“Pass the latent output from this pipeline to the next pipeline to generate an image in a comic book art style” and “need to include the token ‘charliebo artstyle’ in the prompt to use this checkpoint”. ↑b
- 5Comfy Org, ComfyUI, GitHub README, undated, read at source 12 Oct 2026“Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints”. ↑
- 6Hugging Face, IP-Adapter, Diffusers documentation, undated, read at source 12 Oct 2026“IP-Adapter checkpoints are typically ~100MB because they store adapter weights, not a full model.” ↑
- 7Hugging Face, Pickle Scanning, Hub documentation, undated, read at source 12 Oct 2026“Most notably, it is the default format for PyTorch model weights.” … “There are dangerous arbitrary code execution attacks that can be perpetrated when you load a pickle file. We suggest loading models from users and organizations you trust” … “Every time you upload a pytorch_model.bin or any other pickled file, this scan is run.” ↑
- 8Hugging Face, Safetensors, documentation, undated, read at source 12 Oct 2026“Safetensors is a new simple format for storing tensors safely (as opposed to pickle) and that is still fast (zero-copy).” ↑
- 9Hugging Face, Load adapters, Diffusers documentation, undated, as cited on this site’s LoRA page, read at source 1 Oct 2026“It works by inserting new weights into the diffusion model and then only the new weights are trained instead of the entire model. This makes LoRAs faster to train and easier to store.” ↑
- 10Hugging Face, Text-to-image (StableDiffusionPipeline), Diffusers API documentation, undated, read at source 12 Oct 2026“Instantiate a DiffusionPipeline from pretrained pipeline weights saved in the .ckpt or .safetensors format.” ↑
Hugging Face, Pickle Scanning and Safetensors, read at source 12 Oct 2026.
The comparisons and examples on this page (the video-game save point, the anime-style swap) are this site’s own ways of explaining, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK