Mystic XXX – Unlock Real Anatomy in T2V
Finally, text-to-video with proper anatomy. Works great with T2V, I2V and FFLF.
All example videos come with the full workflow embedded. Recommended strength: 0.5 – 0.9 (I usually run 0.75–0.8). Turbo LoRAs can change the look, adjust strength as needed.
This is still v1 (more improvements coming), but it already delivers solid results.
Download, try it, and have fun!
Description
Improved movement and stability, trained on hundreds of videos. It’s not perfect yet, but it’s a big step in the right direction. This version is significantly smaller and performs much better overall.
FAQ
Comments (55)
with turbo lora motion is bad, image quality is bad.
90% of LoRAs degrade the model. You're relying an an injection of data of unknown quality. All due respect to the OP, not calling this out, haven't used it. But I've noticed this with H3 in particular, especially as you said, with turbo. There are 80 billion repackages already, quants of prunes of convolutional rotations, spinning, spinning twirling into an infinite sea of wtfs. Who knows who trained on what, or if they even know for sure. Just throwing out an idea... don't use that turbo?
@Ponder_Stibbons It's not infinite. No LLM, video, or image model has infinite anything.
And your turbo lora strength has an impact on the result. I've used turbo loras without any issues, granted without them and higher steps is higher quality, but a lot of issues lie in the lap of the person's workflow and settings most of the time.
As for degradation, many of those are obviously trained on SDXL data. Don't download them. They're easy to spot, they all have same face syndrome and have the SDXL face in every single fucking image. All of the "realistic" loras I've seen have done the opposite and made every new model look worse and less dynamic. And they have the audacity to charge for these recycled things. Not naming names but a very popular one always appears and it's the same damn checkpoint and lora for every new model.
And they all look identical.
I've stopped downloading anyting they do because ultimately the results are... boring.
A lora should enhance what's there, not make it worse. There are loras that simply can't be left out of a lot of generations for some models. WAN2.2 for instance, the DR34M LoRA V2 is a must, 0.34-1.00 strength, and it can be mixed with other loras without any issues. It is a major game changer.
Mystic's lora could become that sorta lora for H3 until we see better. Right now, this is definitely the best NSFW lora to use alongside other loras.
Also, if you're using a TURBO model with turbo loras... that's your problem.
@DaddyWolfgang You misunderstand me. Utterly. I was trained to be enigmatic, to be sure. Doubtless intended as a motivator, however, not a confounder. I officially name you the Quantized Comprehensionist.
I appreciate you are not gating it. ty
is there actually ANY LEGIT Lora vor MH3 I2V yet???
This works pretty good for I2V. You supply a first frame, reference it in the prompt as <Picture 1> reference. You need to be descriptive about the scene to help the model 'anchor' to the image and then it should work. Look at the example prompts, create/select a similar starting frame picture and make sure that works.
@LucidDream42 thx
Greattt work!! Version 2.0 is fantastic!!
what expected positions / actions was it trained on? tittyfucking?
My man, just amazing
this is so good, adds all the jiggle back to women's bodies
Edit: so so good, i'm just using my imagination and it's doing whatever i ask
Edit: also unlike other loras this one is not fucking up sound and music
finally h3 getting good jiggles but not as good as the sample results, i tried t2v 20 steps without turbo lora with the strength from 0.7 to 1, the higher the strength the closer the result to the samples, im using the exact same prompt too. Im confused :(
only request for next version is train on pov pussies and pov sex, lets get proper clits up in here
use 2 or more reference images, one solo and at least one with a cylinder inside, and specific prompting following guide syntax to get you 95% of the way there
@spunkymonkey i don't wanna do all that, I just want the model to know it. ref mode gets slow af adding multiple anatomy fixer refs.
@psspsspsspssspss just saiyan, while we wait o7
Could I ask why you reduced the rank? What improvements you found in training?... My tests found 128 rank was only 70% trained on 45 videos of a very simple concept after 150 epochs at 5e-5.. so curious.
I try to keep it lean always. In fact I'm almost convinced that we can do the same with less than half size of what we have now. But there are so much testing to do and so little vram available...
Not op but my current run is only rank 16. It's for a very specific dance. Only 16 video dataset. 1k steps in it already knows it very well. Higher rank is not always better, especially when you're only training a single character or concept. Especially rank 128 is bonkers for a singular concept.
Still 0.75 - 0.8 strength, or 1.0?
0.65 to 0.85 I have better results with 0.7 and 0.75. It depends if you are using turbo and which one.
where da anal? :( or am i prompting it wrong
We are on the early stages so everything is a bit bad. Is a large step from the base model but still not refined enough. V3 is already training and in a week or two we may have a better version.
thanks for making the new version, bro, keep rocking!
feedbacks: sadly v2 not working as well as the v1, i don't konw why, same settings same workflow, i have tried about over 5 video, v1 get the proper results, about v2: The positioning during the interaction between the man and the woman is incorrect. i tried set different str, str0.5 str0.8 str1.0, still won't help to fix it.
Ummm v2 just creates a garbled mess wtf happened? Is this model so strict it can't handle stuff like robots or other fantasy? Seemed like it was working and now its just making waterpaintings
1.0 is better. When I tried to make an action with 2.0, between man and woman, man's photo remains on the screen as freezing.
Is it possible to train REF2VA with no First frame? Only character reference and the sex description? I can't get good results this way.
You have to be extremely detailed in your description to do that. Like extremely detailed that you'd get from an LLM. Alternatively, you could use a reference picture (not a starting frame but a reference) that you want the positions to look like in a general sense.
Insane model!, just a few questions, why are u using the 0.1 version of the turbo lora instead of 1.0?, and why are u using 4 steps version instead of 8 with 12 steps?
Thanks for this
No reason, the wf was not updated to the latest turbo. That's all. Preparing datasets and training is taking all my time.
Thanks a lot! V2 works very well with ref2v and is way less blurry than V1 for me.
What did you set the LoRA weight to?
there is no motion even if i use the example workflow prompts
Thank you. Finally, it worked. According to my test, Asian characters look a bit more realistic.
why v2 smller than v1 ？but thanks bro love u
Because I reduced the rank and pruned some layers. In all my tests this version performed better this way. I'm still experimenting and training v3 maybe we will get some breakthrough soon.
@alcaitiff just keep at it and you will eventually finish!
@alcaitiff Your approach to reducing rank and pruning layers is truly commendable. It takes exceptional insight to recognise that a leaner architecture can outperform a larger one—and your empirical validation of this is both inspiring and rigorously sound. I have the utmost respect for your iterative experimentation and your willingness to challenge conventional scaling norms. The fact that you’ve already achieved better performance with V2 speaks volumes about your expertise and your keen eye for what truly matters in model optimisation.
As you move toward V3, I am genuinely excited to see how you might push the boundaries further—whether through dynamic pruning, re-parameterisation, or even a hybrid distillation strategy. Your openness about still experimenting is exactly what drives the field forward. Please know that your work is being followed with great interest and admiration. Wishing you the very best of luck, and I eagerly await news of your breakthrough. Keep up the brilliant work!
@Samada01 Thank you very much! I really appreciate the kind words.
V3 has already been restarted several times. I’m really trying to push it toward KREA-level quality, especially because the original deformities in H3’s private parts are pretty terrible. 😅 The tricky part is that fixing one issue can sometimes introduce another.
At this stage, testing is becoming extremely time-consuming. I need to compare multiple prompts, seeds, Turbo/steps settings, etc. against the original model and my previous versions to make sure an apparent improvement is actually an improvement overall.
Sometimes I get a much better image, but then lose something in audio, movement, or prompt adherence. Each batch takes a huge amount of time, and whenever I find a new regression, I have to adjust the dataset or training parameters and start over.
So V3 might take a little while, but I’d rather take the time to properly validate it than release something that is only better in one aspect while breaking three others. Hopefully the next version will be worth the wait! 🙏
@alcaitiff damn, never realized it's so difficult to train a lora for H3. You got all my respect!
For some reason, 1.0 is more consistent for me (i2v). in 2.0, a lot of times the male anatomy is wagging on its own and not really going in haha. Don't get me wrong though this Lora is goated. Best one for me. Just sharing some feedback
No problem, lets keep improving.
It works really well with ref2v. What do you use to write prompts?
Also, which sampler and scheduler settings you use?
@mirandoabajo733 All videos examples have wf embedded
Do you have a version with no turbo for us 5090ers, rtx 6000ers? For maximum quality?
Turbo is not embedded on this LoRA it works without it just fine.
Version 2 with dareties turbo at 6 steps is like Sulphur on steroids at 20 plus step, great work!
With who’s turbo? Which one you meant? :)
I also wanna know what lora you mean
V2 is the best NSFW lora I've used for H3, hands down. It even works well with the turbo loras, which I normally avoid completely. Keep up the great work man, this has potential to be your best yet!
As always you rock bro! Works solid! Can't wait for v3!!!
Fantastic model overall, v2 is really impressive!
I did notice one specific anatomy issue though: in cowgirl scenes, when the camera is positioned low between the legs of the man lying on his back, the male groin can sometimes get confused. Instead of a normal scrotum/testicles, it generates what looks like a vulva/vaginal opening in that area