A LoRA is a small add-on file that tweaks an AI model to add a specific character, style, outfit, or concept, without replacing the whole model. It layers on top of a checkpoint, so you keep your favorite base model and switch a look on or off. LoRAs are compact, stackable, and easy to use.
If you have seen people share tiny files that add a distinct art style or a recurring character to their images, those are LoRAs. This guide explains what a LoRA is, how it differs from a full model, and, most importantly, how to actually use one you have downloaded. It is written for beginners who want to customize their results without training anything themselves. If you are curious about making your own, this page points you to the full training guide at the end, but the focus here is on using existing LoRAs well.
This article is for adults only and covers concepts intended for people 18 and over.
What a LoRA actually is
LoRA stands for low-rank adaptation, which is a technical way of saying a small, efficient patch for a large model. Instead of retraining an entire multi-gigabyte checkpoint to add one character or style, a LoRA captures just the difference that character or style makes and stores it in a file that is usually only a few dozen to a few hundred megabytes. When you load it alongside your model, it nudges the output toward whatever it was trained on, and when you remove it the model goes back to normal.
The key mental picture is layering. Your checkpoint is the full brain that knows how to make images in general. A LoRA is a thin overlay that says, on top of everything you already know, lean toward this specific thing. Because it is only an overlay, you can keep one base model and swap different LoRAs in and out to get wildly different results, and you can even use several at once. This is why LoRAs are so popular: they give you enormous flexibility for very little disk space, and they let a single good checkpoint serve dozens of different needs.
Another way to think about it is a filter or a costume. The actor, your checkpoint, stays the same. The LoRA is the costume and makeup that turns that actor into a particular character or gives the whole scene a particular style. Swap the costume and the same actor plays a completely different role, which is exactly how a small library of LoRAs multiplies what one model can do.
LoRA versus checkpoint versus embedding
It helps to see how a LoRA sits between the two other ways of customizing output. A full checkpoint replaces the whole model and is the heaviest option. An embedding, sometimes called a textual inversion, is even smaller than a LoRA and teaches the model a single new keyword, but it is much more limited in what it can change. A LoRA sits in the useful middle: small, flexible, and able to change quite a lot without the size of a full model.
| Type | Size | What it changes | Flexibility |
|---|---|---|---|
| Checkpoint | Multiple gigabytes | The entire model and its overall style | Complete, but you swap the whole model |
| LoRA | Tens to hundreds of MB | A specific character, style, outfit, or concept | High, stack several on one base model |
| Embedding | A few KB to a few MB | Mostly one narrow concept tied to a keyword | Low, best for small nudges |
For the full picture on how these pieces fit into the generation process, see how AI image generation works, and to browse ready-made options the best NSFW LoRAs roundup is the place to start. Most people build a small collection of favorite LoRAs over time and reuse them across projects.
How to use a LoRA you downloaded
Using an existing LoRA is a four-step routine, and once you have done it once it takes seconds. First, download the LoRA file, which will be a .safetensors file, from a reputable model site. Second, place it in your software’s LoRA folder, which is usually a folder named loras inside your models directory. Third, add the trigger to your prompt. Most LoRAs invoke with a special tag that references the file, and many also have a trigger word that the model card lists, so you include both to activate the effect. Fourth, set a weight, which controls how strongly the LoRA applies.
That weight is the part beginners overlook, and it is the most important control. A LoRA is applied at a strength you choose, commonly written as a number from 0 to about 1. A weight around 0.6 to 1.0 is a typical range: lower for a subtle influence, higher for a strong one. If the effect is too weak, raise it; if the image looks distorted, burnt, or the LoRA is taking over everything, lower it. This is closely related to prompt weighting, which controls emphasis in a similar way. Always read the LoRA’s model card, because it lists the recommended weight and any trigger words, which saves a lot of trial and error and gets you a good result on the first try.
Stacking multiple LoRAs
One of the best things about LoRAs is that you can use several at once, for example a character LoRA plus a separate style LoRA plus an outfit LoRA. This is called stacking, and it is how people combine a specific character with a specific art direction. The catch is that each LoRA pulls the image in its own direction, and their weights compete for influence. If you stack three at full strength, they often fight and the result looks muddy, burnt, or broken.
The fix is to balance the weights. Lower each one so the total influence stays reasonable, for example running two or three LoRAs at moderate weights rather than all at maximum. Start with the most important one higher and the supporting ones lower, then adjust based on what you see. A character LoRA might sit near the top of its range while a style LoRA rides lower so it flavors the image without overwhelming the character. In node-based tools this balancing is very visual, and the ComfyUI guide shows how the connections work if you want fine control over stacking order and strength.
When to use a LoRA versus a different checkpoint
A common question is whether to add a LoRA or just switch to a different model. The rule of thumb is about scope. If you want a specific character, a particular outfit, or a narrow style added on top of a look you already like, a LoRA is the right tool, because it changes one thing and leaves the rest of your model intact. If you want to change the entire aesthetic, for example moving from anime to photoreal, you want a different checkpoint, because that is a whole-model change and no single LoRA will get you there cleanly.
In short, reach for a LoRA for targeted additions and a checkpoint for the overall foundation. Many users settle on one or two favorite base models and build a small library of LoRAs to vary them, which keeps disk use low and results consistent. If you find yourself stacking many LoRAs just to fake a different overall style, that is usually a sign you actually want a different checkpoint instead.
The kinds of LoRA you will see
Not all LoRAs do the same job, and knowing the rough categories helps you search and combine them. Character LoRAs teach the model one specific person, so you can generate the same original character reliably across many images. Style LoRAs apply an overall art direction, for example a particular illustration look or a film-like grade, without tying you to any one character. Concept LoRAs add a specific idea, pose, or action the base model handles poorly on its own. Clothing and outfit LoRAs add a particular garment or costume. And detail or quality LoRAs subtly sharpen skin, lighting, or texture across the whole image.
The practical upshot is that these categories combine naturally. A character LoRA plus a style LoRA plus an outfit LoRA is a common stack, because each one owns a different layer of the result. Keeping this mental map in mind stops you from reaching for a full checkpoint when a small concept LoRA would have solved the problem, and it makes browsing a model site much faster because you know which category you actually need.
The categories also hint at how strong each LoRA usually wants to be. A character LoRA typically needs a firm weight so the person stays recognizable, while a style or detail LoRA often works better dialed lower, since a light touch flavors the image without taking it over. There is no universal number, and the model card’s suggestion always comes first, but knowing the category gives you a sensible starting instinct before you even run the first test, which shortens the trial-and-error considerably. It also helps you diagnose a bad result: if a stack looks muddy, the style or detail LoRA is the usual suspect to lower, and if the character looks generic, its LoRA is the one to raise.
A simple stacking example
Picture the goal of a consistent original adult character, in a specific painterly art style, wearing a particular jacket. You would load your base checkpoint, then add three LoRAs: the character LoRA near the top of its recommended range so the person stays consistent, the style LoRA a little lower so it flavors the image without erasing the character’s features, and the outfit LoRA lower still, just enough to bring in the jacket. You put each LoRA’s trigger word in the prompt and generate.
If the first result looks muddy, you lower the style LoRA a touch, because a dominant style LoRA is the usual cause of a character losing definition. If the jacket is barely there, you raise the outfit LoRA slightly. You change one weight at a time and regenerate with the seed fixed, so you can see exactly what each adjustment did. Within a few passes the three LoRAs settle into a balance, and you have a repeatable recipe you can reuse. This is the everyday craft of using LoRAs, and it is closely tied to character consistency techniques that keep a face stable across a whole set.
Trigger words and where to find them
Many LoRAs only wake up when a specific keyword, the trigger word, appears in your prompt. Without it, the LoRA can sit loaded and do almost nothing, which is one of the most common reasons a beginner thinks a LoRA is broken. The trigger word is not something you guess; it is listed on the LoRA’s model card by whoever made it, often alongside example prompts you can copy to see the intended effect.
Some LoRAs need no trigger and apply automatically once loaded, while others have several trigger words for different variations, for example different outfits or expressions baked into one file. The rule is simple: always open the model card before your first generation, note the trigger words and the recommended weight, and start from the maker’s example prompt. It removes almost all of the guesswork. If a LoRA still seems weak after you have the trigger word in place, the next thing to check is the weight, and after that the base family, since those three, trigger, weight, and family, account for nearly every case of a LoRA that appears not to work. When you want to emphasize a trigger a little harder, light prompt weighting can help, though raising the LoRA weight itself is usually the cleaner lever.
What to watch for
The biggest trap is a base version mismatch. A LoRA is trained against a specific model family, so an SD 1.5 LoRA will not work properly on an SDXL checkpoint, and vice versa. If a LoRA seems to do nothing or produces garbage, check that its family matches your model first, before anything else. The checkpoint families explainer covers how to tell them apart so you can match correctly.
The second trap is over-high weight. Pushing a LoRA past its comfortable strength fries the image, adding artifacts, color banding, or a burnt look, and it can distort anatomy badly. If your output degrades as you raise the weight, back it off toward the middle of its range. The third is a missing trigger word. Many LoRAs only activate when their specific keyword is in the prompt, so if the effect is absent, confirm you included the trigger the model card lists. A fourth is trusting the source blindly; download from reputable sites and prefer safetensors files. And as always, keep every generated character adult and fictional.
LoRAs in a node workflow
If you use a node-based tool rather than a simple slider interface, LoRAs work the same way conceptually but look a little different. Instead of typing a tag into the prompt, you add a LoRA loader node between your checkpoint and the sampler, and set the strength on that node. Stacking several LoRAs means chaining several loader nodes, each with its own strength, so the balance you would type as weights becomes a visible chain you can see and reorder. This makes it very clear which LoRAs are active and how strong each one is, which is handy when a stack gets complicated.
The tradeoff is a little more setup for a lot more clarity and control, which is why many people who work with heavy LoRA stacks prefer it. If that sounds appealing, the ComfyUI guide walks through building the graph. Whichever interface you use, the underlying rules do not change: match the family, set a sensible weight, include the trigger word, and balance the stack. The node view just shows you those choices instead of hiding them inside a text tag, and it makes debugging a misbehaving stack noticeably faster.
Verdict
A LoRA is a small, stackable add-on that customizes an existing model with a specific character, style, or concept, without the cost of a full checkpoint. To use one, drop it in the loras folder, add its tag and trigger word to your prompt, and set a weight in the rough 0.6 to 1.0 range, adjusting from there. Match the LoRA to your model’s family, keep weights balanced when stacking, and read the model card for the recommended values. Browse the best NSFW LoRAs to get started, and if you eventually want to build your own character or style, the complete LoRA training guide is your next stop.
Frequently asked questions
What is a LoRA in AI image generation?
A LoRA is a small add-on file that modifies an existing model to add a specific character, style, outfit, or concept. It layers on top of your checkpoint rather than replacing it, so you keep your favorite base model and switch a particular look on or off. LoRAs are compact, usually tens to hundreds of megabytes, and can be stacked together.
How is a LoRA different from a checkpoint?
A checkpoint is the entire model, several gigabytes in size, and it sets the overall style. A LoRA is a thin overlay, far smaller, that changes one specific thing on top of that model. Use a checkpoint to define the whole foundation and a LoRA to add a targeted character, outfit, or style without swapping the base model out.
How do I install and use a LoRA?
Download the safetensors file, place it in your software’s loras folder, then add the LoRA’s tag and any trigger word to your prompt, and set a weight. A weight around 0.6 to 1.0 is typical: lower for a subtle effect, higher for a strong one. Always check the model card for the recommended weight and trigger words to activate it.
What weight should I use for a LoRA?
It depends on the LoRA, so the model card’s recommendation comes first. As a general starting point, a weight around 0.6 to 1.0 works for most. If the effect is too weak, raise it; if the image looks distorted or the LoRA dominates everything, lower it. When stacking several, reduce each so their combined influence stays balanced and nothing fries.
Can I use more than one LoRA at once?
Yes, this is called stacking, for example a character LoRA plus a style LoRA. The catch is that each pulls the image in its own direction and their weights compete, so at full strength they can fight and produce muddy results. Balance the weights, keeping the most important higher and the supporting ones lower, then adjust from there.
Why is my LoRA not doing anything?
The two usual causes are a base version mismatch and a missing trigger word. A LoRA trained for SD 1.5 will not work on an SDXL model, so confirm the family matches your checkpoint. Also check the model card for a required trigger keyword, since many LoRAs only activate when that specific word appears in your prompt alongside the tag.
Should I use a LoRA or a different checkpoint?
It comes down to scope. If you want to add a specific character, outfit, or narrow style on top of a look you already like, use a LoRA because it changes one thing. If you want to change the entire aesthetic, such as anime to photoreal, switch checkpoints, because that is a whole-model change no single LoRA achieves cleanly on its own.
Do I need to train a LoRA to use one?
No. There are thousands of ready-made LoRAs on model sites that you can download and use immediately, which is what most people do. Training your own is only necessary when you want a character or style that does not exist yet. If you reach that point, a dedicated training guide walks through the full process from dataset to finished file.
