My take on the cumshot lora.
Trained on 60 videos, Rank 32, Sigmoid, 5400 steps. Not trained on audio - most of the dataset would just be dudes grunting. Videos focused mostly on facials, but it seems to generate well on specific parts.
Samples generated with standard workflow + comfy kitchen, 20 steps, euler/beta. Strength set at .85. Metadata should be included with the videos.
I've had issues with turbo loras creating artifacts around the cum.
A fun little note: The first 2400 steps would just generate still video's of penises.
Step 3000 generates truly unhinged stuff. I have included that as a wildcard. Sometimes it's pink. Sometimes it gives girls big cocks. Sometimes it generates the scene as described but inserts another woman passionately kissing each other. None of these things were in the dataset. While I may not be the audience for this save, I know there is one on this website. So I have included that for you.
And after 3600 it actually does what it's trained on. This save seems to provide the best results with minimal body horror.
Description
FAQ
Comments (64)
Hi, thanks for sharing. At first glance, the cum shots look intense and more realistic. I'm going to test out your Lora and post the results :)
I'm referring specifically to the video I made and converting text to video only
Big up!!
Everyone, this is the one! works great in ref2vid and doesn't even need a penis in frame to work.
any hints on how to get the penis off screen? Thanks in advance.
@bailung938 whenever I used turbo Lora’s it had a knack for inserting them into places they should not be so that may be an issue with keeping them on screen. I haven’t tried this with ref2va yet, but I would use language that involves the phrase out of frame/off screen. Though the Lora is trained at 5400 steps so it may be pretty baked in to the concept in general.
@lt_warf what strength/settings did you use? I find it very jittery in Ref2v
@Sassyphrassy 0.85 with the 8 step turbo Lora as well at1.0. nothing else.
@lt_warf Awesome Ill give it a shot. Thanks for the reply
most realistic yet
SO VERY RELIEVED that we're now back on a video model that can conceptually connect the source of an action with the result. After Wan, LTX felt awful, where fluids would just appear completely unrelated to fluids that were shot out, and they were shot out from random sources to begin with. Wan was SO much better than LTX for physics, and Minimax H3 has Wan beat. LTX felt like the dark ages.
I don't think we're there yet, but we're clearly getting much closer. Cum seems like one of the "last hills" for sure. Seems like a matter of time. Wan does it so well with minimal prompting, where most Minimax stuff needs a lot of prompt support to get good results. (That and Wan looks much better at lower resolutions, but I digress.)
LoRA learnings still seem pretty early, but it's exciting to see everything emerge. Seems like early days still.
@boobkake22 i think it can be done, but you would have to perfectly curate the dataset to the training FPS. I’m no expert, but from my experience the models understand the general concept, but something like this it will hallucinate the origins of the fluid that get skipped in between frames and it seems to want to fill in the details. Wan seems to want to fill them in out of frame. Minimax just adds more quantity to the initial release. Or I’m just an idiot, I dunno.
@boobkake22 Whether it's hard to prompt or not, and whether it's perfect or not, this and what Wan did were pretty good. LTX simply shit the bed at every attempt. There wasn't even a concept of "good enough" during those dark times. I'm just relieved that things are good again. And this is the best yet, even getting the color much closer, instead of white paint.
But sure, there's space between what this is, and perfection. My point is that we are emerging from some very dark times, and I feel so much relief.
@Jellai Yeah, we're in agreement. :) And this can do shockingly specific things you could never dare to do with Wan. Very rad times.
A rope of stickly fluid lands over her eye with a distict string of sperm hanging across her eyelashes.
...actually works.
Greattt work!! High quality and realism!!
Well done. The first lora that can genrate realistic cum on skin.
Work really well. will test the R2V later.
My sample 40s video is on review and will be added later.
The only thing that was not rendered corectly in my short movie, is the erupting phase from the tips of the cock. Even if i used the exact same sentence from one of your sample prompts.
But the cum on the skin is perfect.
Now we have two perfect cumshot loras, CUMOUF for cum in mouth, and yours for facial, and all others things :)
Thank you.
EDIT: Ok, i removed the penis lora in my second 50s video, and this time the cumshot is perfect :).
most of the dataset would just be dudes grunting
lol
Tonight we goon
What resolution were you training on?
384
@alternative_penguin Sorry 384 square pixels right? (Great job btw. One of the best H3 loras I've seen, with the quality. I thought you were training much higher resolutions!)
I never thought I could fall in love with a Penguin. @alternative_penguin
But here we are...
So I did some testing and the Penguin figured out the magical formula. ALL HAIL PENGUIN! ALL HAIL PENGUIN!
It under a month H3 has proven its the king. I'm using the plaguekind workflow and their lora and it's just doing everything correctly and it looks great doing it.
Took a bit of fiddling but I can say for sure that for me it's the best workflow and their lora does a better job and can be ran at higher CFG's and maintain image and audio quality.
The community is really making some great shit.
I have moved all the WAN models off my drive and all the loras. No need now.
How do I keep it from looking like huge gushes of milk?
for the samples I fed an abliterated Gemma model the prompt guide and just ask it to create a T2VA scene that’s xx seconds long. The scene should describe (include basic idea). Use language to enhance realism and textures. Feel free to embellish details to keep the scene erotic/seductive/sultry/whatever mood you want. No music and include if you want it to include cuts or be one continuous shot.
Also using turbo Lora’s seems to increase this issue in some of my testing.
Thanks, I'm using a 4-step accelerated checkpoint; that must be it.
This is the first really good Lora for Minimax, no idea how you managed to make her look so good. I hope you do more.
finally a penis that looks like a penis in T2V.
Amazing work, @alternative_penguin. Highest quality LORA for H3 I've seen so far.
How well does it work with the "Worship It"? Any recommendations on the strengths or was it built on top of it?
I ran a few gens before I realized how turbo loras seem to effect this one by itself. Might be better with i2v or r2v. I'll try again tomorrow.
Absolutely amazing work man. By far the best cum(shot) lora!
could you share you comfy workflow to achieve the videos in your samples mine never look real like that
If you download the videos you should be able to drag and drop them into comfy and it will load the workflow.
Nice work, gotta step up my game ;)
Appreciate your work on the templates and guides! Guys like you got in early and proved tackling these models was possible.
You're both doing great work. Keep going.
nice to see creators working together and not infighting with each other :)
Dang now THIS is praise. You're both incredible creators, thanks to you both for expanding the community resources
Great and versatile.
one of the best!! works weel, looks good!
Works really well but can't use turbos which is a little frustrating. Has anyone found a good strength combination to get food results with turbos?
I ran a bunch of generations last night across a few templates and fresh installs. I can safely say the turbo Lora’s are not causing the pixelation some may be getting. I assume it’s caused by another speed technique, I just can’t figure out which one. I typically generate with a barebones install of comfy optimized to use with pro 6000 instances. I have used quicker templates to run tests during training runs, which is how I noticed the pixelation issue. I like to use those templates and workflows to generate samples because I assume that’s how most of the people downloading this will use it so it helps sets realistic expectations.
I run it with turbos all the time and am getting excellent results. Depending on the turbo lora i usually will not run them beyond 0.75.
@DaddyWolfgang thanks I'll have to try again
This lora works REALLY well, good job!
great lora! One problem is that, the lora seems to be trained with only cumming viewing from the dorsal side of penis, so it tends to switch the penis side towards dorsal side.
That’s 100% of the training. Maybe try starting your prompt with a description of the opening shot pretty vividly. You probably won’t get good results though.
I always get the woman with a penis !
any idea to solve this ?
Prompt for the man out of frame. Character Lora’s can cause this as well.
@alternative_penguin ah, okay, thx. but I found another way, using another mm-model.
@jmadrox57174 Isn’t the open source community great?
Sounds like a feature not a bug!
hands stick on dick, i think this is what should be improved, what ever thanks for good dicks
Check my wan Lora for training tips!
The best cumshot lora for H3 yet. :) However it does seem to lessen the effectiveness of other blowjob/deepthroat loras though. If it's strength is lowered to around 0.30 and below it starts working better with other loras but then the effect of this lora of course also get weaker. :/
Or maybe it's just me?? Anyway if there's a way to fix this in future releases I would very much appreciate it.
Thank you for creating this lora, it really is top-notch.
The Lora’s won’t stack well, even my own. Take a screen shot of your final frame and run it as an I2VA prompt. Maybe running the motion Lora at a low strength will help with the action, but it will always degrade the quality.
@alternative_penguin That's a good tip, thank you. Yes simply extending and changing out the lora and prompt would bypass this. Or maybe using a reference video. I'm already using a reference audio clip to fix the carrot crunching from the other blowjob loras so another motion helper won't hurt I guess. haha. :) :D
And hey don't sweat it man. Minimax H3 is still very new, it's not fair comparing it to Wan 2.x loras since they have had 1-1½ years to get better. :)
Really nice results!! What trainer did you use for?Would love to know—thanks!
AI Toolkit. Rank 32, sigmoid timestep. 60 videos, this uploaded model is step 5400.
Recipe for Turbo users:
- 8 steps spectrum
- larry600 at 0,75 + light2v at 0.45
- then search for this lora:
LMS - A Little More Sharpness
and set it at negative value -0.33One thing i noticed after lot of testing.
stream of semen don't appear at 4 steps, with 4 step, only the cum on the skin appear, the stream are almost totally invisible.
For it to appear in the video, it need at least 8 steps, at 6 steps it start to appear but not fully.
I don't know if this could be fixed for afuture version, so stream of cum appear even at only 4 steps.
I tested with int8 convrot, eros max beta5, redcraft, etc...and also NSFW AIO lora.
This Lora is someth too strong for my setup. Using the strength 0.5 is helpful.