Civitai Models and LoRAs: How to Pick, Read and Use Them

Civitai (civitai.com) is the main community library for Stable Diffusion and Flux models. It is where most checkpoints, LoRAs and embeddings get published, discussed and rated, and almost every example image carries the prompt and settings that made it. Knowing how to read a model card, which file type does what, and how to combine them saves hours of guesswork whether you generate locally or on the site.

Checkpoints, LoRAs and embeddings

Civitai lists several kinds of file and the labels matter, because they are not interchangeable.

• Checkpoint. A complete image model. It decides the overall look, anatomy, lighting behaviour and how prompts are interpreted. You load exactly one checkpoint per generation. Checkpoints are built on a base model family (for example SD 1.5, SDXL, Pony-style SDXL derivatives, or Flux), and everything else you add must match that family.
• LoRA. A small file that modifies a checkpoint to add a style, a character, a garment, a camera look or a concept. You can stack several, each with its own weight. A LoRA trained for SDXL will not work on an SD 1.5 or Flux checkpoint.
• Embedding (textual inversion). A tiny file that adds a new word to the prompt vocabulary. Often used for negative prompts ("bad hands" style embeddings) or a particular face shape. Also family-specific.

The practical rule: choose your base family first, then only browse models that match it. Civitai’s filters let you narrow by base model, so set that filter before you start browsing.

How to read a model card

Every upload has a model card, and the useful information is spread across a few places.

1. Version selector. Many models have several versions. Newer is not always better; read what changed and look at the example images for each.
2. Base model tag. Confirms the family (SD 1.5, SDXL, Flux and so on). Check this before downloading anything.
3. Trigger words. LoRAs and embeddings usually list one or more words that activate them. Without the trigger word, a character or concept LoRA may do very little; a style LoRA often works without one but behaves more strongly with it.
4. Description. The uploader’s notes on recommended weight, sampler, CFG and what the model is good or bad at. This is the single most skipped and most useful part of the page.
5. Example images. Click any image and look for the generation data panel: prompt, negative prompt, sampler, steps, CFG, seed, checkpoint and the LoRAs with their weights. Images uploaded by the creator are the most reliable reference; community posts vary.
6. Licence and permissions. A box near the top shows what the uploader allows: whether you may use generated images commercially, sell the images, run the model as a service, or merge and redistribute it. These are set per model and they differ. Read it, and screenshot it, before you use the model for anything paid.

Using trigger words and weights

A LoRA is loaded with a weight, usually written in the prompt as <lora:filename:0.8> in Automatic1111-style interfaces or set in a loader node in ComfyUI. The weight controls how strongly it alters the checkpoint.

• Start at the recommended weight. If the card does not give one, 0.7 to 0.8 is a sensible first test for most style and character LoRAs.
• Put trigger words early in the prompt. Place them in the first few words, exactly as listed, including any odd spelling or underscores.
• Lower the weight when it takes over. If every image has the same face, pose or colour cast, the LoRA is too strong. Drop it in steps of 0.1.
• Stacking. Two or three LoRAs are usually fine when their total weight stays roughly below 1.5. Combine one style LoRA with one subject LoRA rather than two of the same kind, which tend to fight.

Copying settings from example images

The fastest route to a good first result is to reproduce an example exactly, then change one thing. Open an image, copy the full generation data, and recreate it with the same checkpoint, LoRAs, sampler, steps, CFG, size and seed. If your output matches, you know your setup is correct and every later difference is yours. If it does not match, you are usually missing a LoRA, using a different VAE, or on a different version of the checkpoint.

Keep your own record of what worked. A LoRA name, weight and trigger word combination is easy to lose, and when a model is removed from the site its card goes with it. See tracking prompts across AI tools.

Settings at a glance

To generate without a local setup, see the Civitai on-site generation guide.

Try it

Checkpoint: a photoreal SDXL checkpoint of your choice
LoRA: one film-grain or analogue-photo style LoRA at 0.7, with its trigger word first
Prompt: [trigger word], candid street photograph of a woman in a yellow raincoat crossing a wet road at dusk, neon reflections, 35mm lens, shallow depth of field
Negative prompt: blurry, low quality, deformed hands, text, watermark
Sampler: the creator's recommended sampler, 30 steps, CFG 6, 1024x1024, fixed seed

Generate once, then change only the LoRA weight (0.5, 0.7, 0.9) with the same seed. The three images show you exactly what that LoRA contributes, which is the knowledge you need before mixing it with anything else.