A lora trained on worshiping cocks.
Samples generated with Plaguekind's V6 workflow, all settings were his defaults. Each video should contain the metadata, just drag and drop into comfy.
v0.2 Trained on 2700 steps.
v0.6 Trained on 6200 steps.
For those interested:
Both versions are the same training session. This was attempt number four.
First attempt: I tried training on the non pruned version, but that did not work at all.
Second attempt: trained with Linear timestep. The motion was good, but it just made weird penis’s.
Third attempt: trained with shift. Similar to the second attempt, but better prompt adherence.
This attempt was trained with the sigmoid timestep. I just let the training run all day, 6000 steps may be overkill but what I did notice is you start to get the correct visuals after step 2000. The saves in between generated a lot more body horror. These may not be perfect, but they gave the best results along the 7 sample prompts I threw at each save so I uploaded both.
I’m going to upload a cumshot Lora I trained tomorrow. That one.. was a learning experience.
Description
Trained on 6200 steps.
FAQ
Comments (28)
incredible work!
Interesting - seems like the audio is the thing that benefited from the extra training of v0.6(?). Less of that crunching on apples sound that Minimax H3 seems to like so much :-D
EDIT - just saw that your v0.2 wasn't trained on audio. I'm assuming the higher step training also didn't include audio either - so probably just random chance that the audio came out better on v0.6. I should pay more attention, really.
incredible work. AI is going to ruin the porn industry.
There is no porn. There is only Minimax
I enjoy the logical turns it makes the prudes do.
Fantastic work!! The quality of the physical body contact and the movements is incredible! Looking forward to more of your work!
This works great with prompts lifted from your videos but I couldn't trigger it otherwise, what triggers it? Are there trigger words in your prompts? or the camera work? ima confused.
dataset used was just natural language captions, no trigger word.
Any reason why you deleted my comment? I was praising you for not cherry picking your samples?
I haven’t deleted any comments. I appreciate any praise though!
Edit: I trained this Lora on what I wanted it to do. I thought it did the job well, so I shared it. I can see the parent comment you are talking about mentioning “wonky videos” in my notifications. I did not delete anything. I could care less about negative feedback, I’m just adding to the open source community.
My comment was also deleted without reason when I wrote a comment on another model.
This is really amazing! Faces keep consistent. The penises remain as they are and don't grow into long and thick fleshy cucumbers! Thanks!
Congratulations! Finally, a physically nearly perfect penis and not only that, the other actions works great too! I was about to create a LORA for the cumshots because I'm not satisfied with what's been done so far by others, but if you're already working on it, I'll let it go.
How does this preserve the face better than without any loras? You could make another lora only for face preserving lol
Both of your comments have the same implication: the dataset was likely heavily prompted with very verbose descriptions that cover as much about the scene as possible. If the training data prompts everything, it'll be better at knowing what part of the video is a face or not -- hence preserving faces. If the provided prompts are doing better than your own, it implies the training data either 1) repeats the same phrases too often, but based on this question and the quality of the outputs its more likely 2) its using such unique and verbose prompts for each piece of data that you're having trouble scratching at it.
Frankly you already have the solution, which is use their example prompts. I do this all the time, most of my prompts are just frankesteined versions of examples or uploads. They already have a key, you can just add your teeth.
After the latest Eros release, other models, and this and other Loras, H3 has officially gotten better at what WAN and LTX can do by a huge margin. And it'll just get better. I've freed up over 1TB on my drive.
I'll eat crow, I was completely wrong about what H3 would be. I thought it'd be some highly censored poor copy of LTX and WAN. I couldn't have been further from the truth and I'm glad to be wrong on that.
Character facial and anatomical consistency is incredible, if using reference photos and videos then it's basically 1:1 throughout the video. People having issues need to check their weights, steps, and make sure they don't have the turbo loras cranked to the fucking moon or don't use one at all (I never go over .35 and usually just keep it them at 0.25 or don't use one).
I think it's just the sad part of having to increase steps. Like if the turbo isn't high enough you're definitely needing over 15 steps to actually get the right fidelity. I mean I want to run 20 steps every time, I'd just rather have 2 videos instead of 1. Even if it's slightly blurrier lol.
I deleted everything WAN related as well.
My WAN workflows were prone to crashing a lot, and I was forced to restart comfyUI many times, it was akin to a MO while working with WAN.
H3 hasn't even crashed once. The difference is night and day when compared to WAN. It is much less taxing on the system.
The one thing I will say is that Minimax H3 is clearly not trained on any kind of porn, and not even nudity, because feminine intimate parts are completely wrong, or ignored, without LoRAs. That said, it is by FAR the best video model I have ever tried, and it will likely only get better and less taxing to run from now on (hopefully).
It is not very effective for a small, flaccid penis; it transforms the flaccid penis into an erect state. I really enjoy having a woman take my flaccid penis into her mouth and move it around, and I hope to see more data on this.
It is a little concerning that, although most creators have hit near-perfect baselines with these low ranks, the higher-variety datasets aren't coming through. I think that's at least "Level 2" or 3, though, and it does seem like a handful of creators are at least getting there. This is effectively "new" anatomy, it'll take some time to find the right balance in the prompts and datasets.
This is the case for every lora. Lmk if you found a solution.
@Tvz6 If you have already provided an image of a flaccid penis, you need to describe its shape in great detail and disable any LoRA that includes genitalia; this prevents the original shape from being distorted, though the resulting pose may not be ideal.
hi, thanks for lora. what is SCALE needs??
This is the only LoRA I found that can get rid of those carrot cracking and apple biting bj noises. But unfortunately at strength levels this takes effect, it changes the appearance of the character and the lighting of the scene and camera instructions for me. (REF2VA, no turbo, dasiwa checkpoint v2, 0.6 strength, 25 steps)
These vegetable/fruit noises are such a PITA. Some ppl seem to be able to manage them, but I don't get there.