AI Glossary · model-family

What is a Checkpoint?

Definition
A checkpoint is a snapshot of all the trained weights of an AI model at a specific point in training. For Stable Diffusion, a checkpoint is a single 2–7 GB file containing every parameter needed to generate images. Checkpoints can be fine-tuned from base models to specialize in specific styles, subjects, or aesthetics.

Every AI model is defined by its weights — millions or billions of numerical parameters learned during training. A checkpoint freezes those weights into a portable file that can be shared, downloaded, and loaded by anyone with compatible inference software.

For Stable Diffusion, thousands of community-fine-tuned checkpoints exist — each specializing in anime, photorealism, architectural rendering, or specific brand aesthetics. Unlike LoRAs (which are small adapters), full checkpoints are the entire fine-tuned model.

How it works

Base vs. fine-tuned

'Base' checkpoints (like SD 1.5 or SDXL) are released by the model creators. 'Fine-tuned' checkpoints are trained by the community on additional data, specializing the model for specific styles or subjects.

File formats

Most checkpoints are distributed as .safetensors (safer, faster loading) or the older .ckpt format (can contain arbitrary code, security risk). Always prefer .safetensors.

Types & variants

  • Base checkpoint
    The original released model (SD 1.5, SDXL, SD 3.5).
  • Fine-tuned checkpoint
    Community-trained variant specializing in a style or subject.
  • Merged checkpoint
    A mathematical blend of two or more checkpoints combining their strengths.

Common use cases

  • Loading a photorealism-specialized checkpoint for product photos
  • Switching to an anime checkpoint for stylized character art
  • Sharing custom-trained models as portable files
  • Archiving specific model states for reproducibility

Frequently asked questions

What's the difference between a checkpoint and a LoRA?
A checkpoint is the entire model (2–7 GB). A LoRA is a tiny adapter (100–200 MB) that modifies a checkpoint's behavior. Checkpoints are standalone; LoRAs require a compatible base checkpoint to work.
Should I download .ckpt or .safetensors files?
Always .safetensors. The older .ckpt format uses Python pickle, which can execute arbitrary code on load — a real security risk. .safetensors is cryptographically safe and loads faster.

Further reading

Related terms

What is Stable Diffusion?
Stable Diffusion is an open-source latent diffusion model that generates images from text prompts. R
What is LoRA?
LoRA (Low-Rank Adaptation) is a fine-tuning technique that teaches an AI model a new style, characte
What is Diffusion Models?
Diffusion models are a class of generative AI that learns to reverse a noising process — they start
What is a Checkpoint in AI Models? Explained | Oakgen | Oakgen.ai