For more LoRAs and updates, Join my Discord:
https://discord.gg/ZVWVhT43GW
You can deploy MiniMax on RunPod with my template: https://get.runpod.io/minimax-template
Follow me here so you see the next one.
Recommended settings:
T2V/I2V:
Turbo LoRA - 8-steps version at strength 0.20
Cumshot LoRA - strength 0.9
Sampler - Euler
Scheduler - ddim_uniform (Optionally include the ExtendIntermediateSteps node and add 2 steps starting at sigma 0.85)
R2V:
Turbo LoRA - 4-steps version at strength 0.85
Cumshot LoRA - strength 0.9
Sampler - Euler
Scheduler - beta (Optionally include the ExtendIntermediateSteps node and add 2 steps starting at sigma 0.85)
V0.5:
Expanded dataset (56 videos) with more angles
A thorough description of the cum texture helps a lot!
V0.1:
A video LoRA for the cumshot itself: cum leaving the cock, travelling, and landing on her face/tits/pussy/sister/father in law
This is mostly a demo as I am testing different LoRA training settings, dataset is very slim and limited to 1 POV position, but it does other position as well.
I will get to training another version with a more robust dataset in the future.
Prompting
It was captioned in plain, blunt language, so write prompts the same way:
hmcumshot3, he shoots a big load of thick white cum from his cock in five thick
spurts. She is a petite blonde with her eyes shut and her mouth open. This pov
shot shows her kneeling on a dark wood floor.Related
HMPussy and HMPenis are anatomy adapters. This one is an action and it stacks on top of them.
Description
FAQ
Comments (48)
Thx for pushing things in the right direction!
Thanks for the feedback!
This is the first best cumshot lora for H3, keep it up, but please be careful to not train to model to to create a spray of white paint
Yeah it's tricky lol.
Thanks for the nice feedback.
The dataset here is really slim so it's stuck to mostly a single angle, but I plan to train it again soon.
This is a very very good start, i was starting to loose hope, but between some jiggle loras, this cumshot lora and minimax h3 saying that they are working on fixing their current model (distorted faces etc). I think we can finally put wan 2.2 behind us in a few months. I think this model still has some unnatural stiffness, at times, and fine details can get blocky and pixelated/distorted and face consistency is hit or miss, but Im optimistic it can all get ironed out in the future. Even though training is tougher, ppl are still spittin out loras and turbos at a rapid pace. Cheers!
the first proper H3 Lora! Well well done!
The head of the penis was not properly treated.
Use my Penis lora to correct that
@HearmemanAI damn! at the first glance I read "Use my Penis to correct that" 😂😂😂
Fantastic!!
At the risk of sounding extremely gay: we desperately need an H3 LoRA with a nice-looking penis.
Use the Penis Lora of the same author?
HA! GAYYYYYYY
You just feed a nice penis reference and boom you're done
@bhass Comfy OOM's when trying to use my massive schlong as reference.
@Lora_Addict None of the Penis Loras get all the way, at the moment. Most penises still end up looking like claymation a bit.
@Dasher787 hmm works great for me, but as bhass said, use a reference image and you're good
Honestly, other minimax cumshot loras so far only looked natural from certain angles or poses, but this one is so insanely well done that it's hard to believe it was trained on such a little dataset.(Only doggy style cumshot feels a bit off) Already looks great, but I’m really looking forward to the next version trained on a bigger dataset. Please make one!
Great job on this one. Considering this is just a demo, I'm ecited for the first real release. If you need someone for some testing before, feel free to contact me. Thanks for this fantastic LoRA.
Anyone else having this problem ? The character's voice is actually read the trigger word. "hmcumshot3", is so annoying. I'm using a turbo lora with 6 steps and 1 cfg. Thank youuu!
I've added a "LoRA Trigger words: ..." before the trigger words, and it seems to have helped...
Maybe next time, it could just be trained on action phrases?
Drop the trigger word it will work the same
Have noticed this happening with other trigger words, too. Particularly "hmmotion". The person will say the correct dialogue written in the prompt and then just end with "Hmm... motion". Really annoying lol
@razr112 Yeah, I think with video models that have audio, and that take more of the prompt into account, we may have to ditch contextless keywords, unless defining an actual action that fits within a sentence. It seems like it's a holdover from booru, and when prompts were purely visual.
No need for trigger words in my case.
Thank you guys, it worked! Dropped the trigger word.
Man, after being stuck with LTX for so long, it's nice to get back to a model that can train concepts like physics, with the source/action of something being at all related to the result. Wan could do this back in 2.1, but LTX robbed us of it for a good while.
makes better cum, but literally doubles gen times and messes with prompt and voices. Thanks for the lora though!
Thank you
Cum looks like thick yoghurt or milk product more than actual cum. Lora strength on 1.0
Maybe he's sick???
instead of "white" describe it as "translucent clear off-white viscous". Other models I've had luck with "separated", getting the clear + the off-white chunks. idk too much though, haven't tested a whole lot without also mixing in the MysticXXX lora, it seems to understand cum pretty well
Looks natural to me, some a bit more thick than normal, but otherwise it's okay. If your cum is clear and watery then you need to see a doctor. It shouldn't have the consistency of water nor be clear. Like Makia said, you sometimes need to be very clear with what you want to see so "semi-translucent milky (off)white cum" works in a lot of cases.
Fellas, your cum shouldn't be have the consistency of water and overly clear. Healthy ejaculate is an off-white color and thicker than water. You might just have a vitamin deficiency or something else entirely is wrong.
@makiaeveli I've just used the suggested prompt : 'he shoots a big load of thick white cum from his cock in five thick spurts', maybe the problem is in the 'thick white' description. I mean it's like Greek Yogurt or something.
@sarahbonestone To be fair that is the wording in the provided prompts. My only real suggestion would be to try and reduce those vague kinds of words and replace them with just trigger-word related verbs. I see "eject", maybe "ejaculate". Even though "thick" might be in the training data, it's just trying to explain itself. If you explain it better, it will still retain physics and such from the lora.
@makiaeveli Yep, that's the solution, I've just tried, instead of "white" describe it as "translucent clear off-white viscous", definitely don't use thick white. Drastic changes in end results.
I just replaced my cum description with this from his example prompt and this absolutely solved the yoghurt problem in v.05. Looks amazing now!!
Here is the part I used:
hmcumshot3,his cock pulses and thick white cum shoots out in 5 separate spurts. The texture is explicit: dense, pearlescent, ropey and viscous, not watery, first spurt a heavy glossy rope that holds shape mid-air with a wet sheen, second and third thinner but stringy with long elastic strands webbing from the slit, fourth and fifth more clotted and starchy, creamy curds that bead and cling. It lands hot on her face while it is still moving: first rope across her closed eyelids coating lashes, second across her cheek dripping slow and leaving a translucent trail, third into her open mouth pooling on her tongue, fourth and fifth thick driblets collapsing onto her chin and upper lip forming a white glossy pool that sags and drips down toward her throat. Her expression tightens at impact, eyes squeezed shut, nose scrunched, lips parted letting out a wet surprised laugh, ribcage heaving, shoulders tensing. Her face stays locked in frame throughout, focus locked hard on the nearest wet surface of skin and cum, the background sliver going soft, autofocus hunting once as the first spurt crosses close to the lens.
This is secretly a comedy lora. too much fun.
I'd love tips on how to prevent the over-enthusiasm of this lora, so the penis isn't launching a payload that resembles a bucketful of cum.
Toning down the lora strength doesn't seem to help.
try giving it a shorter shot timeframe, with the next shot (____ stops) or the liek
ok will do, thanks. I did also find that if your location is too distant, then the jets have to become more powerful to reach it. In other words, if your penis is a long way from the top of your target's head and you want it splattered, either specify that he moves closer or he'll turn his gonad pumps up to 5000 to make the distance.
In the initial dataset I scoped to large cumshots on purpose.
V0.5 has it a bit toned down. but still ropes.
I'm looking forward to an update on this, right now it's pretty hard to use.
1. the cum is too thick and viscous right now, you can actually dilute it if you go above the recommended 0.9 but it affects the image a lot so it's not viable.
2. even at the recommended setting of 0.9 the image integrity goes out the window when doing i2v - even at 0.4 mp it will do weird things like change the background.
That being said it's the only option right now.
You got it
@GlowingGuardianGirl I agree it can be a prompting issue, but 0.4mp shouldn't be a problem.
Thanks for the kind words
@GlowingGuardianGirl This is completely irrelevant. Minimax H3 is one of the first models where no matter which resolution you generate, it should look good in terms of prompt adherence, action, etc.
Sure 720p is always better than 360p. What I'm tryna say is that this model is very different from LTX 2.3 and others where if you had a good working prompt on 720p. The moment you bump the resolution to 1080p, your seeds start to look and behave differently creating a completely different result.
Minimax H3 is pretty consistent across each variation of megapixel.
@damiang Wasn't even talking to you, and I strongly disagree with you. As soon as you add a LoRA H3's power can completely colapse and you just have to browse the video gallery to understand it. People break models then whine, so next time just mind your own business.