New ref2va file to download (experimental, no audio trained)
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: 0.2 – 1 (V4 I run at 1). Turbo LoRAs can change the look, adjust strength as needed.
Download, try it, and have fun!
Description
This version is more stable and works better with higher strength. 0.9 gave me great results.
FAQ
Comments (76)
How does it go with anal?
Was there much in the dataset?
Can you please train for 'spitting'? Specifically cum and saliva. Mostly saliva. The details of saliva and the action of the girls spitting/drooling. No one ever does this
Yes yes yes. And cum bubbles/spit bubbles.
I'll take requests for Lora's . My first one for grinding just came out well. DM me if you have something specific you want
@MarisGolding My comment says what I wanted
yep nothing is better than this lora for the moment absolute cinema great job
Text 2 Vid anus confirmed, genius sir!
only cons in r2v: it makes the breasts way too bigger than the reference.
I think the lora is maybe overfitted...
@dizi0n might be a great lora on t2v, but yes, on r2v it destroys the details, as well as have high chance of anatomical error at the weight of 0.7 or above.
@dizi0n Or the dataset is lacking small breasted women, a variety of shapes and sizes. The lora is working great though. looking really good in r2v for me.
Amazing quality, definitely going to be one of the top ones out there!
Version 2 works incredibly just as a motion lora. It also helps the audio a bit when using turbo loras.
thank u
With V3 I am struggling to get quality output compared to V2. In particular audio seems destroyed when using this lora and at high strengths, and it introduces a lot of noise to the video as well. I have tried 0.25 - 0.9 in conjunction with 8-step turbo and fail to get as good as results as V2.
I used ref2va, and version V3 showed significant improvements, with good performance in male motion speed and genital structure.
Detailed settings: fl2va int8 model, V3 Lora strength 0.8, enabled ComfyUI Kitchen Attention (significantly improves generated structures), sampler res_mulitistep + simple 14~20 steps, turbo Lora not used.
seems like you merged too many loras on V3. it fights just about every single lora i stack on top of this other than the speed loras. by itself it's great, especially for anatomy but, actions themselves are pretty generic. can you let us know which lora's you merged to make V3?
emmm, v2 better, v3 make sth difrent and break audio. still thanks for your work.
How good is ref2va at using reference images for identity, scenery, audio, etc.? I mean references that don't already have the pose or action. Good work, by the way.
nice
I found this lora to works the best for ref2va videos without turbo. euler + beta.
I say this because everything I do above 0.5 comes out overexposed and blurry. And cumshots are just streams of white pee
Great anatomy but all videos turn out oversaturated now.
everyone comparing v3 & v2, just use both. it is better to have these loras, both have their pros and cons.
@TaiLong loading unnecessary LoRAs only leads to more VRAM usage and slower generation times. If it works, it works, but it is not ideal to use more than you need.
@zkyt well , nothing is perfect.
It's just feedback to let him know that there was a regression in v3 for a lot of people. That's what the comments section is for isn't it?
@blaqua No problem, all feedback is good. V4 is on the oven and I'm listening you.
@zkyt @blaqua ok, i take it back.. v2 is really better ..i tried both & compared them.
@TaiLong All good bro, glad we all agree and OP is working on a fix already
Detailed settings: fl2va int8 model, V3 LoRA strength 0.8, res_multistep + simple 8-step sampling; Turbo LoRA was not used. V3 exhibits some distortion/warping, whereas V2 is relatively more stable.
你们都用的什么模型啊？
fl和ref混合模型
minimax_h3_hybrid_fl2va_ref2va_b30-49-int8。plaguekind那个工作流是我目前看到唯一可行的SLA加速工作流，这个模型是它自带的。它是我测试下来所有方案中，目前生成速度最快的一个（1MP+13s的视频在3090平台速度大约70s/it）
Author, will your next model optimize the structure of the vagina? I really want to make a video of playing with the clitoris, but in fact, it's a bit difficult
By the way, the lora is extremely powerful
For this I suggest you to create the image first and use I2V, my KREA or Ideogram can create good base images. V4 may have a better T2V but is still to early to say
@alcaitiff
Thank you for your idea. I will give it a try. your lora is very useful. I'm looking forward to a better version. Keep up the good work!
I don't know if it's this lora or the turbo lora I'm using, but the video turns bright white 1 second in with the audio completely distorted (fl2va I2V).
Can you please donwload any of my videos drop on comfyui and test with my config?
any trigger words?
yeah was going to ask this. i think for other mystic models you have to use specific trigger words like bl0wj0b. curious if that's needed for these loras
It doesn't require any trigger words; once connected, the effect is immediate! And I also really hate anything that uses trigger words.
works awesome in ref2va, not sure why but setting the turbo lora at 2.0 really improved results (makes it look burned tho)
If it makes it look 'burned' then how is improving results. Also, what is burned?
@Piraterage23 check the videos I posted, it's the turbo look or too much cfg look. If you crank up a turbo lora to 2.0 you'll see it. But strangely with r2v and mystic it's not too bad, and the anatomy/prompt adherence dramatically improved. I have no explanation, you can check my workflow embedded to try it.
I'd drop it to 0.25 or 0.35 and you'll get excellent realistic results.
AWESOME!!!!!!!!!!!!!!!!!
any good base lines to start at using turbo?
Not sure what I'm doing wrong with v3 and Ref2VA. It comes out completely distorted. 20 steps, res_multistep, simple, 672x1216
You should switch to the I2V workflow.
grab a hybrid model, it fixes ref2va on so many levels. 20-49 is pretty good in my experience
https://huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models
@acedelgado143 Agreed. The new Eros beta2 hyrbrid is VERY good as well at just 20 steps.
v2 works great with ref2v but v3 doesnt work at all with ref2v
@chaiwala008215 Yeah, I did an A/B test with ref2v and when using v3 with the exact same other settings the woman in my gen would no longer connect her mouth for a bj and just mimic the action next to the penis.
@geek137 keep using v2 same as me on r2v until we get an updated version for r2v
THIS IS THE LORA FOR NSFW, the man has done it again, my bro fixed all major issued with the previous one, and delivered even more.
Thanks, LoRA V4 still baking. Metrics look solid so far. H3 remains stubborn as hell about converging cleanly.
what type of generations do you do?
v3, 0.6 strength makes the character's voice crack and creates a soft-ish layer on top (the issue is more obvious with non-realistic video)
And if you use it with Turbo LoRA, it amplifies everything.
I understand why it happens, but still it's very annoying
Turbo lora really isn't all that much faster and you can turn that lora down and still get a gain in speed. Make sure you're using the larger 4step turbo lora (1.2GB) and put the strength down to around 0.25-0.35 and set to 20 steps. A 20 second video takes around 3 minutes.
@DaddyWolfgang I only use v3 LoRA, and the issues are already visible, though lowering its strength does somewhat reduce.
You can A/B test with a character talking about something with an accent (not necessarily in a xxx scene), baseline vs Mystic LoRA and the voice change is very noticeable.
this post isn't actually a complaint
Edit: I checked meta data and yours has
"ss_audio_loss = -0.5
"audio is whatever, ship it."
rank = ∞/∞
1.21e9 gigawatts
..."
:)
V2 is way better. V3 fries style and text render.
Your Excellency, your V2 version of xxx is the best H3 LoRA I have ever used. It performs amazingly when combined with various LoRAs for enhanced motions and cinematic camera movements.
Unfortunately, when running V3 alongside other LoRAs, I’m encountering heavy audio noise, blurry results, and horrendous visual glitches with badly twisted limbs.
Regardless, thank you for your selfless hard work. The world is better for having creators like you. Can’t wait for V4! Keep it up!
Huge congratulations on releasing V3!
I’ve been following your journey and know how much sweat and tears went into this. Balancing image quality, audio, motion, and prompt adherence is no joke – and you actually pulled it off!
The fact that you restarted multiple times just to avoid trading one issue for another shows your dedication to real quality, not just quick releases.
If V3 is even close to KREA-level, it’s already a massive win for the whole community. Can’t wait to test it out and see those “deformities” finally gone! 😄
Respect for your patience and thoroughness – this is what makes a great creator. Keep pushing! 🔥
After testing (I2V), V3 performs well in terms of animation, but the image becomes blurry, and the characters' faces become unrecognizable. Currently, V2 is the better option.
Finally, something that works! V3 at 0.6 strength works really well! Just do enough steps and it'll clear up any voice issues!
Testing V3. getting pretty good results. you cooked with this one :) .
Any tips / recommendations on prompting? Not sure what kind of language it accepts. All i've tested so far is prompts from examples. How creative can you be with this? For example, how would you prompt a hj
Which merges did you do with the latest?
No merges just training
@alcaitiff come on bro.... seriously....
Does this work well with the latest Eros Model? Thanks for your work, amazing!
I'v trying it with 10Eros_Max_h3_TURBO-hybrid_beta3_int8_convrot.safetensors with ref2vid but I get bad results so far. still testing
Do you have any plans to make a v4 [KREA 2] Mystic XXX Lora as it's definitely my favourite nsfw LORA along with realism engine v3. You could maybe integrate it with lodestones/Kroma and or silveroxides/Kroma-LoRA-Extract from Huggingface.
You are awesome thank you so much for all of this great work !
Ooooh my, these look amazing! Thanks for your work.
Details
Files
MysticXXX_MMH3-V3.safetensors
Mirrors
MysticXXX_MMH3-V3.safetensors
MysticXXX_MMH3-V3.safetensors
MysticXXX_MMH3-V3.safetensors
MysticXXX_MMH3-V3.safetensors
H3-Mystic-XXX-TFL2V-V3.safetensors
MysticXXX_MMH3-V3.safetensors
MysticXXX_MMH3-V3.safetensors
MysticXXX_MMH3-V3.safetensors
carolstermanv4.safetensors
MysticXXX_MMH3-V3.safetensors
MysticXXX_MMH3-V3.safetensors
mist2.safetensors
MysticXXX_MMH3-V3.safetensors