Checkpoint Models & Architecture

A saved state of a trained AI model containing all its learned weights. In Stable Diffusion, checkpoints are the main model files that determine the overall style and capabilities.

A checkpoint (also called a model) is the complete trained state of an AI image generator. Different checkpoints produce vastly different results.

File Formats

  • .safetensors: Modern, secure format (recommended)
  • .ckpt: Legacy format, can contain executable code

Types of Checkpoints

Base Models

Official releases from model creators:

  • SD 1.5 (~4GB)
  • SDXL (~6.5GB)
  • SD 3 (~8GB+)

Fine-tuned Models

Trained on specific content:

  • Realistic (photorealistic focus)
  • Anime (Japanese animation style)
  • Artistic (painting, illustration)
  • Specialized (architecture, products, etc.)
NameBaseStyle
Realistic VisionSD 1.5Photorealistic
DreamShaperSD 1.5Fantasy/versatile
Anything V5SD 1.5Anime
Juggernaut XLSDXLPhotorealistic
Pony DiffusionSDXLStylized

Using Checkpoints

  1. Download from Civitai or Hugging Face
  2. Place in models/checkpoints folder
  3. Select in your UI
  4. Different checkpoints may need different settings

Merging

Checkpoints can be merged to combine styles, creating new models with mixed characteristics.