AI Guide – Episode 4 – LoRAs

Hello, Hexxet here,

This episode will deal with LoRAs and therefore should also be not Software dependent. (So you can use this knowledge with Forge, InvokeAI, or whatever) :). Please note that I will talk about LoRAs in a more general sense and give you some ideas on how/when to use them rather than pointing out specific LoRAs.

What is a LoRA?

A LoRA stands for Low-Rank Adaptions and you can look at it as a mini model that can’t work on its own but can be added to existing models to extend their knowledge of certain aspects like body shapes, clothing, breast size…

Where do I get LoRAs and how do I add them to my Forginstallation?

On civitai use the Filter by Model Type and choose LoRA (and/or LyCORIS). You might have to further restrict your search to a certain type of base Model because SDXL LoRA’s might not work very well for Pony and vice versa. (However, you may very well try. By now there are quite a lot of hybrid models out there that can handle either base model LoRAs).

Once you have found a LoRA you want to try, download it and put it into the following folder of your Forge installation: webui > models >Lora

Tipp: You can create folders in your Lora Folder such that you can group/manage your Loras in a better way if you tend to amass a lot of them! 🙂

By the way, there is also something called “LyCORIS”, I don’t know what’s the technical difference. But they seem to be more powerful/impactful than LoRAs but you can use them exactly the same way as LORAs. (Put them into the same folder and use them like LoRAs from the software. At least in Forge ^^)

How to add a LoRA to your generation?

Either on Txt2img or Img2Img in Forge you will find a Lora tab, see the image below:

Here you can find all the Loras you have added to your LoRA folder. To add a Lora to your current prompt you can simply click on the LoRA and it will be added to your prompt like f.e. this: <lora:cuteness-xl1:1>

Note: The “1” after the “:” indicates the strength of the LoRA. This can range from 0 to 0.5 to 8. (It probably can go up to any number). Usually, you want to keep a LoRA’s strength between 0.3 and 1. However, this is very LoRA dependent and hopefully, the civitai info to your Lora will contain a strength range. (If so you should probably use that ^^).

The higher you set the strength the more likely it is that your generation will include the feature you tried to add with this Lora. But the higher the strength the more the LoRA will also influence the overall style of your generation. (Which can be especially shitty if you try to do a hyperrealistic generation but want to incorporate an anime-trained LoRA for a pose or something ^^).

So, yeah, you can try to tweak a LoRA with the strength parameter, but in the end, some LoRAs simply won’t yield the desired results in combination with some models. That’s life! ^^. (In general, it’s best to find LoRAs whose preview images already look like the images you try to generate from the overall style).

Furthermore, some (not all) LoRAs need an activation text. If so, you usually find those in the detail section of civitAi under “Trigger Words”:

In this case, you should add this trigger word to your prompt whenever you add this LoRA and want to see its effect. In other cases, there might be multiple trigger words (because a LoRA might do more than one thing and you can control what to apply by adding only the trigger words you want). If a LoRA needs trigger words and you don’t add them, a LoRA might not have the desired effect.

Managing your LoRAs:

In the LoRA tab (of Forge) you can click on the right top corner of each of your LoRAs to open the LoRA’s detail tab:

Aside from some technical information at the top, a LoRA’s information tab might contain a section with training dataset tags. By clicking on these you can add them to your “activation text” below. You may treat these tags like the trigger words I mentioned above. (Though some LoRA’s don’t have this section but still need trigger words. In this case, I advise to copy-paste the trigger words to the Activation text and save.)

Furthermore, you can use the Preferred weight slider to set a default weight that is used when you add the LoRA.

Also, there is a negative prompt textbox. This adds text to your negative prompt. (I never used that one).

If you set activation text and/or a default strength Forge will always add the LoRA to your prompt with that strength and the specified activation text. (So you should definitely do that, especially when LoRA’s need trigger words that you use often)

Also, you can use the button at the bottom “Replace preview” to add the image currently generated to this LoRA as a preview image. (This might help to better identify your LoRAs if you have many of them :)).

What to use LoRAs for?

In this section, I want to list some ideas of what LoRAs can do for you. However, there are probably much more applications so the list won’t be complete:

• Boobs: Most Checkpoints can already do big breasts, and you can regulate them pretty well by simply stating (small/big/huge/gigantic). However, especially the more realistic checkpoints have an upper ceiling here. With LoRA’s I guess the sky’s the limit in terms of size. Also, a checkpoint usually only supports one type of breast. You may be able to tweak the breasts by adding “round” or “saggy” but if you want perfectly round-shaped fake tits or more natural-looking breasts your checkpoint might not be able to do it. You can use a LoRA for that.
• Age: You can influence the age of your characters by writing woman, teen, MILF, young or old, but different checkpoints handle these words very differently. And some checkpoints are only able or good at generating persons in a certain age category. There are so-called age-sliders that let you age up or down your characters more easily by simply adjusting the strength of the LoRA.
• Style: You might want to add a cinematic style to your SDXL generation or you want a different “drawing” syle for your anime generation. You can achieve that by adding a style LoRA. (I’m usually pretty happy with my checkpoints so I’m not using this a lot)
• Details: There are LoRAs with the aim of enhancing details. This may include background items, or more detailed skin, hair, breasts, faces, whatever. If you are not content with the grade of details in your generation, these might help.
• Poses: Hopefully your checkpoint can already do basic poses like standing/kneeling/laying etc. And hopefully, it does understand sex if you are trying to do NSFW generations. But if you want a special pose, a LoRA might be the only way to go to get that threesome rolling. There are LoRAs for double blowjobs, paizuri, and much more.
• Clothes: You might want to add special clothes to your generation. (Like fetish clothes unknown by your checkpoint). You may also just add a normal outfit via a LoRA to have better control of what your character wears if you want to stay consistent over multiple generations. (Note: I usually try to keep my clothes consistent with my prompt. But using a LoRA sounds like a good idea too :))
• Transformations: There is a lot out there. Like dollyfication, robots, Furry, hell there is even a LoRA to turn a mouth into a pussy! Usually Pony and Illustrious are more likely to succeed with these LoRA’s as they already understand the concept of Transformations to some extent. (And most of these transformations are also trained on drawings/comics/anime)
• Concepts: Like MC, spiral eyes, the viewer holding a phone/remote/pendant, or walking in on somebody on the toilet. The concept LoRAs try to capture situations and add them to your checkpoint.
• Lactation/Squirting/Peeing: Most Pony/Illustrious checkpoints can do this already. Most realistic checkpoints have problems with that. A LoRA might help here.
• Tattoos and piercings: Most Checkpoints can do it, but if you want special Tattoos or just a lot of them, a LoRA is probably the way to go.
• Actresses: There are LoRAs of actresses and pornstars available as well. So if you want to generate images that look like your favorite starlet it is possible to use such LoRAs. However, I think all of these LoRAs are tagged as “Don’t use them for lewd stuff” – so don’t :).
• Speed: You think LoRAs can only influence how your generated image looks? Wrong! They can even influence generation speed. (Of course, these LoRAs exert quite some influence over your checkpoint, but you might actually like the resulting style and they can speed up your generation by a lot. (Like 5-10 times!)

There are probably a lot more applications for LoRAs, but that’s all that comes to my mind right now.

LoRAs I like

In the following I want to give you some links to LoRA’s I found helpful or cool. Please note that I only use LoRAs that are free for commercial use for my generation. So, there might very well be better LoRA’s out there that do not fulfill this requirement for me. (There are especially a few in the MC/hypno department I know about :C)

Sliders: I love sliders! And there are sliders for a lot of stuff. In general if you are dealing with something that can get gradually more or less, I would advise looking for a slider. Heck, just search civitai for “Slider” and you will be amazed :).

Here’s a pretty good age slider Imo: Age Mimimeter: https://civitai.com/models/981211/age-mimimeter

Please be careful when using the negative part of the slider in NSFW creations as results can get too young quickly!

Furryfication-Slider: https://civitai.com/models/484952/furryfication-slider-pony

Not really what I am into, but I tried it out and was amazed that this worked out as a sort of slider. If you are into furries this might be something for you. (Note: I think this only works with Pony or Illustrious).

Triboob: https://civitai.com/models/427366/triboob-ponyxl

Three Boobs anybody? I think this one works with Pony and SDXL. But if you have trouble with SDXL there seems to be another made for SDXL as well: https://civitai.com/models/835156/triboob-sdxl

Bombshell: https://civitai.com/models/654712/bombshell-xl-big-busty-boobs-pony

busty bimbo girls? Maybe try this Lora. Also available in SDXL: https://civitai.com/models/147525/bombshell-xl-big-busty-boobs

Concept Slutty CLothes: https://civitai.com/models/523290/concept-slutty-clothes

Comes with a lot of tags that help with making your girls look slutty. (Many newer checkpoints probably can do it already with the right words, but this kinda helped me a lot :))

Details:

isPretty: https://civitai.com/models/287548/is-pretty-sliders-ntcaixyz

detailed: https://civitai.com/models/229213/extremely-detailed-no-trigger-slidersntcaixyz

clear: https://civitai.com/models/279601/clear-deblur-sliders-ntcaixyz

Polyhedron SDXL skin details: https://civitai.com/models/135160/polyhedronall-sdxl-10-skin-hands-eyes-mf

If you want to go for hyper-realistic SDXL creations the Polyhedron Lora is really nice!! (So you don’t need that one if you are going for the speed LoRA (see below)).

Dolls:

1: https://civitai.com/models/102781/inflatable-dolls

2: https://civitai.com/models/140797/inflatable-dolls-xl

3: https://civitai.com/models/309600/pony-xl-inflatable-dolls

4: https://civitai.com/models/887716/adult-love-doll-or-style-or-pony-lora

5: https://civitai.com/models/865379/living-doll

6: https://civitai.com/models/418234/pony-diffusion-ball-jointed-dolls

7: https://civitai.com/models/502111/silicone-doll-001

I’m not really sold on any one of these and I always try them or a combination of them when I’m dipping back into doll generations. The most important keywords for dolls are shiny, glossy skin, and, if you like, “doll joints”. Some checkpoints can already create doll-like girls with just these words, but if you want a certain style or more dollification try out some of the above LoRAs :).

Speed up your Generations

Speed: SDXL Lightning LORA: https://civitai.com/models/350450?modelVersionId=391999

I’m using the 8steps Lora. But 2 and 4 steps are said to also work nicely. With this Lora, you can reduce steps to 8 (I’m usual doing 10 or 12 though). You most likely will need to downscale your CFG to something between 1 and 2 (depending on your Checkpoint). Also, you will want to use Highres Fix for this Lora… I’m mostly using this for hyper-realistic results and I’m amazed by the results. DPM++ 2S a can take a lot of time, especially with high-resolution fixes, but with this LoRA generation time is bearable :).

Even if you don’t feel the need of speeding up your generation, I would advise trying out this LoRA. It’s producing some pretty good results with most checkpoints.

Thanks for reading!

That’s all that comes to my mind when it comes to LoRAs. I hope I could give you a certain overview over how to use them and which LoRAs I like. If you have a LoRA that deals with Transformation or MC, or you deem otherwise super helpful/cool let me know 🙂