Plain-English explanation
When people say they "downloaded a model" for Stable Diffusion, ComfyUI, or a local chatbot, what they usually downloaded is a checkpoint file. It's the base thing the software loads before it can do anything: pick a checkpoint, and you've picked the fundamental style, capabilities, and quality ceiling of everything you generate afterward. Everything else (a LoRA, a prompt, a seed) adjusts or steers that base checkpoint; it doesn't replace it.
The word comes from training: during a long training run, engineers periodically save the model's current state to disk so training can resume if it's interrupted, or so different points in training can be compared and released. Each saved snapshot is a checkpoint. In everyday use, though, "checkpoint" has come to mean any distributable model file you load into a generation tool, whether it's the first release of a base model, a merge of several models, or the end result of a full fine-tuning run like DreamBooth.
How it works
Technically, a checkpoint file contains the model's weights (the numerical parameters learned during training) serialized into a file format the software can read back into memory. For diffusion image models, that typically means the weights for the UNet (or transformer) that does the denoising, plus often the text encoder and sometimes the VAE, all bundled together. For local language models, it's the weights for the transformer layers.
Checkpoints are commonly stored in one of a few formats: older PyTorch .ckpt or .pt files (which use Python's pickle format internally), or the newer SafeTensors format, which stores the same kind of data more safely. The format is about how the file is packaged, not what technique produced the weights inside it: a checkpoint can be an untouched base model release, a community merge of several models, or the output of a full fine-tune.
Adult-AI use
This is the file people choose when they browse a model-sharing site like Civitai looking for a particular art style, a specific NSFW-capable base, or an uncensored model for local chat. "Which checkpoint are you using?" is one of the most common questions in local-generation communities, because the checkpoint sets the baseline: how well it renders anatomy, what styles it defaults to, whether it's been fine-tuned to be more permissive with explicit content than the original release. A common real-world setup for a custom AI character is a downloaded checkpoint as the base, plus one or more small character LoRAs layered on top to lock in a specific look.
Common checkpoint types
- Base checkpoint. The original release of a model straight from its developers (e.g., a base Stable Diffusion or Llama release).
- Fine-tuned checkpoint. A base model that's been further trained on a specific dataset or style and re-saved as its own full checkpoint.
- Merged checkpoint. Two or more checkpoints blended together by averaging or combining their weights, common in community model-sharing to combine strengths of different models.
- Pruned or optimized checkpoint. A variant with unnecessary training-only data stripped out to reduce file size for distribution.
How it differs from model weights, LoRA, and fine-tuning
Model weights is the general, underlying concept: the learned numbers themselves. A checkpoint is the specific saved file that packages those weights so a tool can load them. A LoRA is not a checkpoint on its own: it's a small delta file meant to be applied on top of a checkpoint, not loaded by itself. Fine-tuning is the activity that produces new or updated weights; a checkpoint is what that activity gets saved into once it's done.
Limitations and misconceptions
- "Checkpoint" doesn't tell you how it was made. The same word covers an original release, a full fine-tune, and a community merge; you generally can't tell which just from the file.
- Bigger isn't automatically better. File size roughly tracks parameter count and precision, not quality; a well-tuned smaller checkpoint can outperform a larger, poorly-tuned one for a given use case.
- Checkpoints from unofficial sources carry real risk. Older pickle-based
.ckpt/.ptfiles can execute arbitrary code when loaded, which is part of why the field has moved toward the safer SafeTensors format for distribution.
Technical context
The linked dossiers show why a checkpoint’s base family determines compatible VAEs, controls and adapters.