Michael Limberger

Need me? Email mike@limberger.ca

The model shop

CivitAI is civitai.com: the biggest collection of Stable Diffusion models, embeddings, and LoRAs out there. Think of it like an app store for this craft.

Three kinds of files

Checkpoints are the actual models. These are the engines that make images. Download these first. Usually 2 to 7 GB, .safetensors format.

LoRAs (Low-Rank Adaptations) are small add-on files that tweak how a model behaves. Want a model to nail a specific style or look? There is probably a LoRA for that. Way smaller than checkpoints.

Embeddings (also called textual inversions) are trained concepts you can call up with a keyword. Smallest files of the bunch.

An embedding here is a tiny trained concept sitting in a small file. You invoke it by typing a trigger word. Same English word as the vectors in retrieval systems. Different job.

Reading a model page

Every model has a page with the stuff you need to know.

Base model is what it is built on: SDXL, Illustrious, Pony, and so on. Pay attention: it matters.

Trigger words are special tags that activate what the model was trained to do. Like VERY AWA, COMICBOOK STYLE for Bridgetoons.

Recommended settings are CFG scale, sampler, steps, resolution. These are not suggestions. They make a real difference.

Sample images show you what the model actually produces. Version history shows updates and changes over time.

Fair warning

Not every model on CivitAI works with every setup. Most are built and tested in A1111, ComfyUI, or Forge. They might act different elsewhere.

Read the model page before downloading. The creator's settings are there for a reason.