Civitai is the largest community hub for Stable Diffusion models, hosting tens of thousands of checkpoints, LoRAs, textual inversions, and other assets created by the open-source image generation community. If you have ever spent an hour searching for a specific art style, character look, or quality enhancement for your generations, Civitai is where you find it. This guide explains the different model types, how to download and organize them for ComfyUI and Automatic1111, how to manage mature content settings, and how to evaluate community uploads safely.

Understanding Model Types

Civitai hosts several distinct asset types, and confusing them is a common beginner mistake.

Checkpoints

A checkpoint is a complete model — all the weights required to generate images from scratch. These files are large, typically 2–7 GB in .safetensors format (or the older .ckpt format, which you should avoid due to the arbitrary code execution risk in pickle-based .ckpt files).

Checkpoints define the fundamental aesthetic and capability of your generations. Examples include:

Realistic Vision v6 — photorealistic people and scenes
DreamShaper XL — stylized, painterly images
Flux.1 Dev — the current state-of-the-art open model for quality

LoRA (Low-Rank Adaptation)

A LoRA is a small delta file (typically 50–500 MB) that modifies an existing checkpoint’s behavior without replacing it. LoRAs are trained to add specific styles, characters, objects, or concepts. You load a LoRA on top of a base checkpoint:

A “Film Noir” LoRA adds high-contrast black-and-white aesthetic to any compatible base
A “SDXL Detail Enhancer” LoRA improves fine texture rendering
Character LoRAs teach the model to generate a consistent person or design

Textual Inversion (Embeddings)

A textual inversion (also called an embedding) is a tiny file — often just kilobytes — that defines a new token for the text encoder. You include the token name in your prompt (embedding:my-style) to invoke it. These are the most limited but also the most portable asset type. They work only with models that share the same text encoder architecture.

VAEs

A VAE (Variational Autoencoder) handles the final decode step from latent space to pixel space. Swapping the VAE can fix washed-out colors or soft details. vae-ft-mse-840000-ema-pruned is the standard fix for many SDXL models with color issues.

Downloading Models from Civitai

The Civitai website at civitai.com requires a free account to download most models. The download flow:

Find a model using the search or browse filters.
Check the base model tag (SD 1.5, SDXL, Flux, etc.) — this must match your local setup.
Read the Trigger Words section on the model page. Many LoRAs require specific words in your prompt to activate.
Click Download on the version you want — models often have multiple training runs.

Using the Civitai API for Programmatic Downloads

# Download with wget using your API key (found in Account Settings)
wget -O my-model.safetensors \
  "https://civitai.com/api/download/models/MODEL_VERSION_ID" \
  --header="Authorization: Bearer YOUR_CIVITAI_API_KEY"

The model version ID is the number in the URL when you are on a specific model version page.

Organizing Models for Automatic1111 (A1111)

Automatic1111 (AUTOMATIC1111/stable-diffusion-webui on GitHub) uses a specific directory structure:

stable-diffusion-webui/
├── models/
│   ├── Stable-diffusion/     ← Checkpoints go here
│   │   └── dreamshaper_xl.safetensors
│   ├── Lora/                 ← LoRA files go here
│   │   └── film_noir_v2.safetensors
│   ├── embeddings/           ← Textual inversions go here
│   │   └── bad_hands.pt
│   └── VAE/                  ← VAE files go here
│       └── vae-ft-mse-840000-ema-pruned.safetensors

After placing files, click the orange Refresh button in the A1111 web UI to detect new models without restarting the server.

Using a LoRA in a prompt:

a portrait of a woman, <lora:film_noir_v2:0.8>, dramatic lighting

The number after the colon (0.8) is the LoRA weight — higher values apply the effect more strongly. Values between 0.6 and 1.0 are typical.

Organizing Models for ComfyUI

ComfyUI uses a similar structure under its models/ directory:

ComfyUI/
└── models/
    ├── checkpoints/
    ├── loras/
    ├── embeddings/
    └── vae/

In ComfyUI, add a LoRA Loader node between your checkpoint loader and your CLIPTextEncode nodes. The LoRA loader accepts the model and CLIP connections and outputs modified versions. Stack multiple LoRA loader nodes to apply several LoRAs simultaneously.

ComfyUI Manager for Easy Downloads

Install ComfyUI Manager (ltdrdata/ComfyUI-Manager on GitHub). It adds a Manager button to the UI that lets you install custom nodes, but also includes a Model Manager tab where you can search Civitai and download directly into the correct directory.

Managing NSFW Content Filters

Civitai’s default filter hides explicit content. To keep it that way, do nothing — the default settings are family-safe. If you are an adult user who wants to see mature models, there is an age verification and opt-in process in Account Settings.

Important: Many NSFW models on Civitai are trained on datasets of questionable provenance. Regardless of your personal preferences on mature content, be aware that:

Model provenance is not always disclosed
Community ratings and comments are your best signal for model quality and safety
.safetensors format is required — never run .ckpt files from untrusted sources, as they can execute arbitrary Python code when loaded

Community Safety Tips

Evaluating a Model Before Downloading

Check the download count and likes. Models with thousands of downloads and positive community feedback are lower risk than new uploads with no engagement.
Read recent comments. Users frequently report issues like embedded watermarks, training data problems, or model quirks that the creator did not document.
Check the base model version. A LoRA trained for SDXL will produce garbage results if you load it with an SD 1.5 checkpoint.
Review the license. Many Civitai models are licensed for personal use only. If you plan to use generated images commercially, verify the model’s license explicitly allows it.

Scanning Downloads

Before running any model file:

# Install the safety scanner
pip install safety-checker

# Or use the official Hugging Face pickle scanner for .ckpt files
pip install picklescan
picklescan -p suspicious-model.ckpt

For .safetensors files, the format itself prevents code execution — it only stores tensor data. Prefer .safetensors always.

Finding the Best Models

The Civitai homepage surfaces trending models, but a few search strategies work better for specific needs:

Filter by Base Model: SDXL or Flux to see modern, higher-quality options
Sort by Most Downloaded within a category to find community-validated models
Use the Images tab on any model page to see what real users have generated — these are more representative than the creator’s cherry-picked examples

Civitai is an extraordinary resource when you learn to navigate it critically. The community has produced models that rival or exceed commercial services in specific niches, and new releases arrive daily. Spend time reading model cards, checking trigger words, and organizing your local model library — it pays off in dramatically better generation results.