Checkpoint
What is Checkpoint?
A checkpoint is a saved version of a trained AI model that you can load and use ( or continue training from ) without starting over.
At a glance
- Also known as
- Model weightsModel snapshotSaved model
- Used for
- Saving training progressSharing pre-trained modelsFine-tuningStyle-specific generation
- Common tools
- Stable diffusionComfyUIAutomatic1111Hugging face
- Related terms
- LoRAFine-tuningDiffusion modelModel weightsBase model
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
How it compares
A checkpoint is a complete saved model containing all learned weights, while a LoRA is a small add-on file that modifies a base checkpoint's behaviour without replacing it. Checkpoints define the core capability and aesthetic; LoRAs refine or redirect it.
Think of it like…
A checkpoint is like a save file in a video game: it captures exactly where you are so you can return to that point later, share it with someone else, or continue from there rather than playing through from the beginning again. Just as different save files represent different points in a game's progress, different checkpoints represent different stages or specialisations of a model's training.
Pro tip
When building an AI image or video workflow, choose your base checkpoint first: it sets the visual DNA of everything that follows. A photorealistic checkpoint paired with a cinematic LoRA will generally produce better results than trying to prompt a stylised illustration checkpoint toward realism.
Types and variations
- Checkpoints vary considerably depending on where they are saved in the training process.
- A base checkpoint represents a fully trained foundational model with broad general capability, while a fine-tuned checkpoint has been further trained on a specific dataset to specialise its output.
- Merged checkpoints combine the weights of two or more models into a single file, blending their styles or capabilities.
- In some pipelines, partial checkpoints or delta checkpoints store only the changes from a base model, reducing file size.
- EMA (Exponential Moving Average) checkpoints save a smoothed version of the weights that often produces more stable and consistent output than the raw training checkpoint.
Ready to make your first scene in Morphic?
Try MorphicCommon use cases
- Checkpoints are used whenever a pre-trained AI model needs to be loaded for inference, shared with other users, or used as a starting point for further training.
- In image generation workflows, creators select checkpoints to define the visual style of their output: for example, choosing a photorealistic checkpoint for product renders or an illustrated checkpoint for concept art.
- In video generation pipelines, checkpoints underpin the base model used by the tool.
- Fine-tuners use existing checkpoints as starting points to train on custom datasets, producing specialised models for specific characters, environments, or aesthetics.
Ready to create?
Direct scenes, design characters, and ship full films
All-in-one AI creative platform with simple, transparent pricing, no speed throttles, and an infinite Canvas for max creativity.
FAQs
A checkpoint is a file containing all the learned weights of a trained AI model. Loading a checkpoint gives you a ready-to-use model that can generate images or video without any additional training.
A checkpoint is a complete model file containing all parameters. A LoRA is a small supplementary file that adjusts or adds to an existing checkpoint's behaviour. You always need a checkpoint to run generation; LoRAs are optional additions.
Yes: sharing checkpoints is common in the open-source AI community. Sites like Hugging Face and Civitai host thousands of community-trained checkpoints available for download, though you should always check the licence terms before using one in commercial projects.
Checkpoint files store all the numerical weights of a model, which can number in the billions for large models. Even compressed formats like safetensors can result in files ranging from several gigabytes for image models to hundreds of gigabytes for large language models.
Fine-tuning from a checkpoint means taking a pre-trained model's saved weights as a starting point and continuing to train on new data. This is far more efficient than training from scratch because the model already has broad knowledge and only needs to specialise.
Common formats include.ckpt (the original PyTorch checkpoint format),.safetensors (a safer and faster alternative widely used in the Stable Diffusion ecosystem), and.pt or.pth files. The safetensors format is now generally preferred for sharing due to security and speed advantages.
The concept applies across most deep learning-based tools, but not all tools expose checkpoints directly to users. Consumer-facing platforms like Runway or Kling abstract away the model selection, whereas open-source tools like ComfyUI allow users to load specific checkpoint files directly.
