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
First version use str about 0.75 or 0.8
FAQ
Comments (78)
Have you tested with anime? Will this change the artstyle?
Yes it works! For T2V you may need to reduce str, the dataset only contains realistic images, but all are captioned as so.
@alcaitiff I mostly use I2V or Ref2v. My first try is generating right now :)
I see the motion is a bit floaty in your examples. should I pair it with, HMNSFW for example, to prevent that?
@KiraNugget I did not tried combining with other loras but it should work yes. This dataset has images and videos but my first focus was to make better anatomy. The next version will focus in better motion and after that we can start adding different positions and concepts.
@alcaitiff Nice, I'm sold! Still testing with lora mixing, i'm doing a long stitched video right now. I will metion this lora as a resource when it's done! Cheers!
Finally! Thank you for your service.
Hell yeah, thanks OP. I F5'd at the right time.
nobody can say you didn't care. hell yeah my weekend is gonna fuckin' rock biiiaaatch
Is the dataset purely plastic surgery style breasts?
No, we have a lot of realistic saggy breasts too. But this model likes fake breasts, I'll address this on the next version. Thanks for the feedback.
I shouldn't have opened Civ before work now this is going to be on my mind all shift. Thanks for your work!!
Thanks a lot, man — really appreciate you!
Once your shift is over there’ll be some fun waiting. Just keep expectations low for now… it’s only the first version and we still have a good amount of ground to cover before it gets really good.
Thank you for your contribution!
Thank you
thanks bro, u are the best! penis and vagina improved, the sex action also became more decent, btw love your krea2 models, keep going
Thanks man! That was the plan from the start, anatomy first, movement later. But if the model gods smile on us, we’ll get there.
This is one of the best LoRAs to date.
Thanks man
please v2.0
What do you want on v2 @luisa_caotica ?
Does this LoRA work with R2V model too? If not, when can we expect to support it?
Yes ! ! Will this work with REF2VA as well?
*Update* It appears to work! Just need to reduce strength to 0.5
It was not tested with REF2VA model, but it may work. If you test it please report back. If it don't work I'll create another version for it.
It seems to be working.
@aurelius great! Can you post some example?
I only have this lora and a turbo lora, in ref2v with 3 images (girl+dildo+background)
i'm using a preview node and I see thet it keeps generating extra "ghost" limbs.
Wow!! Fantastic work!! Thanks and congrats!! This greatly expands the possibilities for both the fl2v model and the ref2v model. ;)
Thank you very much
I see now more lora trainers support ref2a, would love a ref2a version (might take more work on the prompts before training maybe). Its very powerful!!!
Well first, thanks for making a more comprehensive AIO. Second, this lora seems to produce inconsistent framerates in the output when used. What i mean by that is its the first lora ive come across for H3 where multiple duplicate frames appear in the 24 fps outputs, no matter the samplers/scheduler combo, duration, resolution i use. I was mostly making 16;9 portrait and that was the first i used the lora and noticed the issue. Depending on the lora strength there are more or less of them, resulting in more severe stuttering video.
So i spent way too long debugging this and I think it comes down to the Sol-Attn node you use in the workflow. For whatever reason that node blends out the framerate to a less noticeable degree. I dont use that node, so i couldn't gen a video that was smooth like i have been doing.
I was able to load your workflows, which pretty much used the same models as mine outside of a different turbo lora and larger text encoder (same int8 pruned model) and the outputs i made were the same as yours, so it wasnt some system issue. But if i disabled the Sol-Attn node, it introduced stuttering into the video outputs you provided. Basically if you go frame by frame you either see duplicate frames or the motion dramatically change in an inconsistent way. Like its stretching a 16-20 fps video to 24.
Hopefully in future versions this can be improved. Maybe your training data has videos that arent all native 24 fps? I cant really use it as is right now because i dont use the Sol attn (it increases vram use in most cases and degrades quality slightly)
Thanks for the huge feedback! I'll do some tests on this later.
any chance you can train one for ref2v?
People reported that it works on ref2v with very low str like 0.2-0.3 But yes we can train it for ref2v. My plan is to improve movement on this one first. Anatomy is far from perfect too.
Can you add some futanari or shemale content in 2.0? 1.0 is pretty nice for sure, thank you for your work!
Too gay
You didn't use any of the penis or pussy helper loras for your videos? Your lora just makes natural looking genitalia on it's own?
Yes, all videos have wf embedded you can see the config. But no other lora than turbo
It's still far from perfect, but still way better than any other NSFW LoRA so far. It doesn't ruin image quality or prompt adherence, and it actually helps with NSFW motion. That said, it still struggles with quite a few things. Penises can look weird at certain angles or end up absurdly long like a snake. I also noticed the same FPS drops and choppy playback another commenter mentioned.
The LoRA is definitely very strong, so lowering the strength to around 0.4-0.5 worked best for me. It really improved prompt adherence for style and movement, which is a huge plus. Looking forward to the next version!
AN IMPORTANT TIP: To further improve the penis shape in Minimax H3, add this penis LoRA ( https://civitai.red/models/2846693/penishelper-h3-t2v?modelVersionId=3214199 ) for fl2v or ref2v. It helps A LOT! I get great results with the LoRA strength set to 0.4
link, lora and account seem to be gone. any chance some can re-upload it?
@denolim465778 That's strange; it was online until yesterday... If I find a re-up, I'll post the link here.
Another tip, you can add a closeup image in ref2v as an image reference (try to match the pose you are going for) and make it a <Subject> in your prompt.
@denolim465778 I block any one that posts fucked up gay shit, so I often find that I'll block lora makers too... But you're right. This guy seems to be gone.
@LucidDream42 Yes, that's true; I forgot to mention that option as well. That works really well too—it turns out perfectly.
@LucidDream42 thank you. Yes, i tried it once, like this is the "digg" or "peenees" of <subject 2> in <picture 2>, but the result was still a exaggerated big deformed one. was it my prompting, do i need to describe it further?
It's work AMAZING with power 0.2-0.3 on MiniMax H3 Hybrid (fl2va base + ref2va reference pathway) with references + turbo ema600
Are you also using the hybrid custom node? Do you have workflow.json?
same Strength with Turbo LoRA also?
@sar Someone has made the hybrid model, the custom node is no longer necessary.
https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models/tree/main
it may not be perfect now still very good, it has better jiggle and penetration concept... looking forward for future updates !! 😊
Edit : i highly under-estimated this lora .... its sooo cool that it only need correct prompt and can do everything (maybe)...
i seem to be getting a hybrid form of ballsack and vagina. i will experiment some more
Same :(
I tested it on the INT8 Pruned version, and this LoRA seems to reduce the visual quality of the fluid. I’m wondering if there’s any way to optimize it.
use hmcumshot lora
no h3 is not good at fluid by default
@TaiLong I'll try, thanks
works nicely on 4-step turbo lora!
Which lora are you using and which sampler/scheduler combo, steps? I found mystic-xxx only to work with the same setup in the example prompts. Even down to using lightx2v 0.1. 1.0 already gives bad results.
@lost_moon the 600 whatever turbo lora, forget where i downloaded it. at 1.0 strength. normal euler and simple.
今晚试试
Works great so far. I've tried other AIO loras for minimax h3 but could never get it right. I've tested on i2v so far and i'm amazed honestly. Thank you for this :) . Can't wait to see the improvements
the dirty king is back!!!
I'm using a weight of 0.75, and there are harsh, piercing moans. They only disappear completely when I lower it to 0.2. Is there any solution to this?
is there a discord? also anyone using this with r2v?
been working amazing with ref2V. Using basic ComgyUi workflow with turbo 600ema, at 0.4 strength but can be adjusted based on preferences.
bullshitting. It doesn't work, the references are broken with this lora.
@Psy_pmp not really. What resolution are your images? It literally works with the basic ref2v workflow.
Proper anatomy?
The Lora is very strong, I use 0.2, otherwise I would get unrealistic monster penis and crazy motion. I think overall it is better than HMNSFW AIO 2 at the time I am writing this.
I guess it's going to take like a couple of months until we get good nsfw loras for H3. The ones we have now add some effect but the motions and details are just... not goonable, I guess.
always the case but fret not - will be great I think
What I noticed with this LoRA is that it really needs the extra steps (as visible in your workflow) to find its legs. Without a LoRA, and with the resulting restrictions; I can get nice render in 8-10. This one wants the 16. Great work and thank you
@alcaitiff I'm having lots of fun with it. Seems pretty stable with the Turbo. Maybe a little overcooked at 1.0 in a few places. Trying a few at .8 and comparing. Overall I like it.
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.
After testing in diverse flf2va situations - like smashcuts, different sex positions, pov/non-pov, hardcore and softcore, with/without turbo, long video (up to 30s) - I can confirm that this lora is wonderfully consistent. Unfortunately though, it seems to mess up ref2va - so currently the solution is to smash cut fom an image of the subject - or use more advanced trick.
Thanks a lot - really great work and keep it up @alcaitiff .
Thank you very much. I'm working hard on the v2.