WARNING: Only works with the Sulphur LTX 2.3 model
Trigger word: PENISLORA (put it at the front of the prompt)
Stroking: "The man is stroking his erect penis. The penis is visible from the side"
Blowjob: "The woman is giving a deepthroat blow job to the man's erect penis. The penis is visible from the front"
Trained on 215 clips of penises from the front view and the side. With no audio.
Prompt "The penis is visible from the side" for side view of penis head.
Prompt "The penis is visible from the front" for the front view of the penis head.
It is not trained on POV view of the penis at all. It has no training data on penetration or cumming.It has blow job and deepthroat in the training data as well. But its mainly for giving details to the penis and stroking. It also can be prompted without stroking.
This is trained on the NSFW LTX dev merge of LTX 2.3. It will not work on the regular model. I have only tested it with the full dev model not the lora version. Strength 1 works great, us the example workflows provided by a kind gentleman. Swap out the LTX base model for the sulpher dev model and put the lora on both passes.
I trained it up to 20k steps, but it had an amazing result at 4.5k steps. Previous attempts on the BASE model of LTX took 20-30k steps for a worse result. Please train NSFW on this model from now on. I trained it on rank 48, and for some reason it trained very fast. 1.3 s/it. This lora was basically done in an hour or so. I may test other checkpoints later, but 4.5k steps seems good enough for now. It's trained on the default layers for t2v, not tested in i2v. I could do a i2v version maybe if there is enough interest.
Description
FAQ
Comments (15)
I'm seeing multiple loras trained on Sulpher. I saw that Sulpher 2 and Eros 1.0 were released at the same time. Why did you choose Sulpher and not Eros? Trying to figure out why Sulpher is getting all the loras.
Eros is a sulphur checkpoint too. Sulphur is the actual base model and is full strength for training.
Eros is also better for i2v is my understanding
sulphur is the baseline for nsfw, eros is a finetune of it for better img2vid if i understand
@ElGordoAI targeted version for i2v. They're the same data/model but integrated in different ways.
i would choose base LTX at all for training, so is compatible with everythig..
but this is just me
@Agent_Smth training penis or other nsfw stuff on base model cost a lot, i imagine sulphur can reduce training time ..
@Agent_Smth this is also true too. Base loras work just fine with both. Training for sulphur you are looking to target or improve a certain concept inside sulphur but you have to use it a lot to find how and what it needs. Usually you're gonna want to go for 768 clips to add detail instead.
I should try training with Sulpher too and see if the results are better.
@JaimyB it trained the penis concept almost x10 faster and it did a better job.
sulphur has a strong bias wich make it almost impossible to prompt certain things in t2v. i use it as a lora at 50% strenght max, with other loras on top, and i think is the way to go to get more balanced and way cleaner output but also make loras trained on sulphur hard to use or useless at all. you dont need to make a perfect lora that works on base ltx, just train on it i guess. even if sucks by itself then will work on both ltx (with other loras) and sulphur
@Agent_Smth Yes, but certain concepts are SO in Sulphur's wheelhouse that it doesn't really make a ton of sense to aim it at vanilla LTX. Sulphur/Eros is the first time I've felt like LTX could possibly handle anything approaching Wan in consistency, and even physics. I wouldn't want to use Sulphur/Eros as my general model, of course, but certain topics are so at home there, and you'd really only want to do those topics in those models.
will this handle details like size / skin tone of penis?
The training data has a variety of sizes and some different skin tones (though not so much variety) but its not captioned or trained for such (circumsized only as well). I think the base model may have that info though.
Can you please share your dataset