I've been trying several NSFW Loras on realistic models, even trained a bunch of them from models like UBPR, and I still get awful results when they are used on realistic models like epicRealism. Thoughts and advice?
4 Answers
First, make sure your LORA really learn something.
If you did something wrong, your lora maybe not work, even it comes out.
This is known as overfitting or underfitting.
SDXL is more harder than SD15.
You have to give a really great database that AI can learn and of course, a all right train steps is necessary.
Talking about incompatibility between models and LoRAs.
As psoft said, LoRAs trained on anime models won't be compatible with realistic models, but, it can be. For this you need to train LoRA with both captions, natural language and booru tags at the same time. This would lead to better compatibility.
Second, LoRAs are trained with a Clip Skip parameter, for training realistic - 1, for training anime - 2. Both of them will work worse with opposite Clip Skip parameter on generation process in your WebUI. So, if LoRA is trained on Clip Skip 2 - don't forget to set it in your WebUI as well.
Third. Some people training LoRAs on models they prefer or just using. For example - LoRA trained on UBPR. This can lead to incompatibility with other models. You need to use model for training, that as simple as possible. The great example is SD 1.5. You don't need your training model to be hyper-detailed or hyper-realistic. Most models are 95-99% the same (on layers and data level) as basic SD 1.5 model, so, training anything on this model will be the best solution for better compatibility.
But, there is another way to train LoRAs. You can train any LoRA on model you're using. Of course, this is the best way to train Adapts (and the way it meant to be), in this case you should get best possible results, but of course, compatibility with other models can be worse, as i said before.
the results of how loras look depends on which base model/checkpoint was used for the training. training on an anime model (NAI lineage) won't work as well on realistic models outside of the NAI lineage. this is the case because the lora changes certain weights of the base model, which may be different between lineages of models.
the solution could be to train the same dataset on both an anime and a realistic base model.
Beyond this the thing is if they're asking about OTHERS loras on NSFW models these replies still matter but the question of the hour is:
WHICH NSFW models?
Are you using Hi Res Fix?
Are you using A Detailer?
Is this SDXL?
SD 2.1 and 2.0 had nerfed NSFW features, so largely that one doesn't make sense.
And A LOT OF people dont' always use standardized training, I rarely personally do NSFW, ut the few times i have i've tried to stick to NAI and SD 1.5.
Especially for realistic loras on my tick? I've tried to make them CROSS COMPATIBLE - in the idea that you could use it on anything.
UPRM is an extremely hard to classify model, and isn't possibly known dont quote me on this to be severely lora compatible.
That has A SPECIFIC GOAL in mind, rather than an NSFW capable realistic model.
That's maybe there it is?
If you're looking into training be sure TO DO check psoft and everyon'es answer's above.
Just that if you're still struggling to get LORAS to work on NSFW models, it largely depends on what loras, and what models.
EXAMPLE:
A SFW innocent chibi model will not work on an NSFW model for what you might be trying to generate.
Aka: Don't try making NSFW/porn on UPRM of FFXIV lalafell, it just won't work.