All examples are t2v.
This is not only i2v Lora. i2v or using the Ref model would be an "easy" job. The idea here is to teach the correct male anatomy for the base model.
It should perform even better with i2v if the anatomy is already present, but hopefully with this lora it should not be needed anymore.
V2:
Start with strength 0.9 – Works from 0.4-1.5
This was trained A LOT more steps at 2048 resolution, read the model details. I decided to change the model name because it doesn't make sense anymore. This should be a "fix" for the whole male anatomy.
This might not be the final version, but I don't consider this an experimental version anymore.
I like to use the Realism people lora from 0.2 - 0.6 to get rid of the plastic look the model has, specially at higher resolution. But the Realism LoRA can make weird anatomy so you will need to experiment.
Limitations:
It still cannot control cut/uncut and barely show foreskin.
not that many penis diversities. Glans are very well converged now, but the penises can all look the same. Size difference works.
frenulum is barely there.
depending on the pose the anus can still look misplaced, but changing strength and seed helps
motion: it was not trained on motion. If you want sex and cum perfect motion, even though this lora can do it as you can see in the examples, you better try to use it with another lora.
woman: untested. It might work but it was not trained on it. Since MM H3 is a really good prompt follower and learns pretty well.
V1:
This is a first experimental version. Trained with my experimental implementation of Musubi in my GUI. https://github.com/diodiogod/musubi-tuner_simple_GUI
After proper quality control technique (inspired by Ostris AiToolkit), it maintains quality very well.
I intend to keep training it. It was trained at first with just still images at 256x. Why? Since Klein training first on 256 has been positively good. I'm doing a continuation training on 1024 for details now and maybe I'll try 2048.
I trained 10 epochs, each is 1.629 steps. But I choose epoch 8 as it had the best asshole positioning.
Penis shape is still shit. But it's getting there. If you use a i2v for first frame with a penis it should be great.
No trigger word, just reference a man from behind or showing ass
Some supporting trigger words used on the dataset:
glans, large penis, small penis, medium-size penis, wide open anus gap, closed tight anus gap, penis upside-down, scrotum, testicles, low-hanging testicles, hairy, shaved Description
This is a long session of refinements with penis anatomy focus + some other concepts: oral, fuck, cum, penis tip urethra opening, ass touching, anal gap, special "sports". The idea was to refine penis but keep the concept of anus still there.
The result IMO is a good male anatomy, finally. That is why I'm changing the name. This is not just ass anatomy and penis from behind anymore. It still sometimes misses, the opening and glans will sometimes come out wrong. The frenulum is still not there yet, and large foreskin neither. But this is miles better than anything we got from the past loras.
There was a problem along the way. After so many epochs and so many steps (over 50k) the quality of the image started to suffer. I think this is mainly my dataset fault. I fixed it (hopefully) by doing block analysis and chopping some blocks. And as a consequence the lora is also smaller in size. If anyone wants to play with blocks and save their lora try my Realtime Lora fork here for ComfyUI: https://github.com/shootthesound/comfyUI-Realtime-Lora
For reference the blocks preset remerged was: 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 1.30, 1.30, 1.30, 1.00, 1.00, 1.00, 0.50, 0.50, 0.50, 0.50, 0.50, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 1.00, 1.00, 1.00, 1.00
Training lineage: Male_Ass-MinMax-v2.02
Cumulative saved training: 42.807 steps · 33 effective epochs
1. Male_Ass-1024-MinMax-v1.03
epoch 8 · 13.032 local steps
2. Male_Ass-1024-MinMax-v1.04
epoch 2 · 3.258 local steps
3. Male_Ass-2048-MinMax-v1.04-cont
epoch 2 · 3.258 local steps
4. Male_Ass-MinMax-v2
epoch 9 · 6.147 local steps
5. Male_Ass-MinMax-v2.01
epoch 7 · 9.982 local steps
6. Male_Ass-MinMax-v2.02
final saved checkpoint: epoch 5 · 7.130 local steps
FAQ
Comments (38)
This one looks better so far. Im just waiting on an actually good cum and fluid physics lora this model is really terrible at it be default. Even Wan is much better. Lame needing buzz tho crypto is annoying af.
Well this took a week or more to train +40k steps and all... the previous versions are free and this will get out of ea in a week. I normally release all of my loras for free, but this took too much of my time. I hope this works well for you, please post any results on the gallery. Thanks!
And for cum, I did some experimentations by using this + HMCumshot - Cumshot LoRA for MiniMax H3 - v0.1 | Hailuo H3 by MiniMax LoRA | Civitai at 0.4-0.6 and I think it might work well
WAN took a long time to get right too. And H3 can be very good with cum. I've had great results that are better than WAN. Give it time my dude, it hasn't even been out 2 weeks yet.
why tf the pizza coming out his ass 😭😭😭
Would you consider adding some regularisation to your dataset for your next training run, perhaps X% nude women, clothed women, men and women together, etc, non sex and sex acts, it could be much more robust and generalise, less likely to futanari everyone?
He means, can you do lora for female anatomy? Because others tried, but results - nah (I believe it's because models still new and it will took time to know how to train for them, but seems to be author way ahead, must be really good dataset).
I don't know why he would need to do that, considering how overtrained every single model is on female subjects. That would lead to MORE instance of that, if anything. You have plenty of female-oriented LoRAs on this site. Use any one of those instead of trying to co-opt one of the only things we have.
@hushicho Yep yep.
I'm not willing to make a general finetune model right now. This is a specialized model on male anatomy. Sure it would help to have regularization and women anatomy as well, but it would be a lot more trouble and a lot harder to converge on males if I did that.
Also is the LoRA really "futunaring" all the women? Have you tried it? I don't think the lora is that overfit on males, to be honest.
It's funny how you guys apparently don't test the model. I was testing this hilarious prompt I found from other user, and it didn't futanari any of the women... it performed better than any of the other penis loras... the women had no hint of a penis in their vaginas that i could see... https://civitai.red/posts/30689118
Such a great model, makes me wish i were gay so could enjoy it.
You can use this for straight content; Loras don't have to just be applied to the overall concept the Lora was created under. And it's a "male anatomy" lora which I'm not sure why you associate it with gay only.
@DaddyWolfgang I'm not sure why he'd use it enough to know that it's "a great model", if he don't like lookin at dudes. Lol.. Someone's in the closet.. X-D
@DaddyWolfgang I don't know, when i apply this lora i get lots of testicles. Testicles that are plump and round. I'd rather see a freshly groomed and neglected vaginas.
I am completely unable to recreate OP's samples with this lora... Maybe it is me. Maybe it is my workflow.... but I get nothing but mutilations.
Are you using a turbo lora? If so that is almost 100% your issue.
@DaddyWolfgang Tried with and without turbo loras.... body horror all around...
@kermitfrog1202 post a result so I can load and check. All the videos have the workflow, you can test them yourself
Are you using a cache node? H3 Sla node? Or aggressive attention? I've noticed they all add noise and issues.
well, it is your workflow... turn easy cache off, turn refine prompt off https://civitai.red/images/140548288
You can use Spectrum, I've tested with it and I didn't see any large degradation with it
You say that you've ran into problems due to so many epochs and refinements, why not take your most refined dataset and make a second version of the lora instead of continuing to build on top of the same one?
In my experience with image loras, the foreskin is the hardest part to get, since most of the male training is done using cut(mutilated) cocks..
Side note: I honestly think that's a crime.. His body, His choice, it should be..
but yea.. the model always wants to turn the foreskin into a big bulbous cockhead..
Phimosis is another that's even more difficult, cuz then the model is wondering 'what happened to the cockhead??', and it has no idea what the chunk of skin on the tip is, or how it's supposed to operate.
This should be easier if you're using video clips to train though, cuz then you can use closeup clips of the foreskin actually being rolled back and forward in real time. at that point, it's just getting the model to understand that the dome is on the top, the urethra on the bottom, and not produce upside-down cocks..
One question, does this lean heavy into hairy gaping ass, or can it do twink ass as well?
Because that would take a lot more time. The LoRA was still learning and converging, not getting overcooked. Starting again would be a waste in my opinion. The anatomy was getting better at each epoch. The overall image quality (blur, jpg artifacts, things like that) was starting to be learned as well. But block analysis pretty much fixed it, I think. Of course a better dataset would be preferable. But that would mean starting over again.
Yes a video dataset would be ideal for foreskin for sure.
IDK, I have a bias towards hairy guys. But there were plenty of shaved twinks as well on the dataset. But probably not as many as hairy guys.
Wow this works perfect for fisting females best I've ever seen, just lower strength to 0.65 and add a pussy Lora (HMPussy - Pussy/Anus LoRA for MiniMax H3) at strength 1. Thank you so much for this Lora and thank you MiniMax H3
It's really good. it can do anything o/ !!!
@diogod dude this model if absolutely insane!! I fucking love it, you have fixed the issue with any issues with inserting absolutely anything!! I'm having so much fun with this, it works perfect for females, (pussy penetration) i think you need to release this again and explain it works flawlessly for female penetration especially for pussy, its a little tricky for anus but does work with the right seed, you only need to set the lora strength at 0.65 and use a pussy/anus Lora
thank you so much though Bro, its so much fun to use this lora, you can actually get the characters to insert anything your imagination can dream up, im so impressed. MiniMax H3 is the ultimate model now thanks to your input
To be very honest this is probably the merits of H3 as a base model and the right lora combination you found. I'm very glad to hear my LoRA is not overfit on males though!
And the best thing to make everyone know it can be used like this is for you to post your generations on the gallery with the prompt and parameters, so others can replicate it.
Thanks!
Which pussy/anus lora do you use to prevent testicles from appearing? If you don't mind
Wow this works perfect for fisting females best I've ever seen, just lower Lora strength to 0.65 and add a pussy Lora (HMPussy - Pussy/Anus LoRA for MiniMax H3) at strength 1. Thank you so much for this Lora and thank you MiniMax H3
Having a lot of fun with this LoRA, thank you so much!
That’s absolutely brilliant,
The Lora is amazing at helping to guide anal penetration.
Thanks! I'm glad it helps! Post some results if you wish on the gallery!
I've added a post to the gallery with a bracketed set of tests of the same prompt and seed with this LoRA at various strengths: 1.00, 0.90, 0.75, 0.50, 0.30, 0.00 (control). Turbo "Larry" 600 step LoRA, 8 steps, full BF16 models.
Penis LoRA Comparison Part 1
Penis LoRA Comparison Part 2
thanks for testing it
@diogod Thanks for making it. It's defintiely one of the best performers.
https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/tree/main
minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors
https://huggingface.co/QrusherZA/H3_Turbo_ComfyUI/tree/main
minimax_h3_turbo_v4_step600_ema_comfy_pruned.safetensors
Safe
4-8step，It achieves excellent results and is very fast. I hope you can release a next version with more different penises.