MiniMax Hailuo H3 NSFW Unlocked
Very excited to finally release our new MiniMax Hailuo H3 NSFW Unlocked LoRA!
After the massive response to our Flux & Klein NSFW models, we took the same high-standard training methodology and brought it to the most capable video model in AI right now: MiniMax H3.
This model completely removes generation guardrails, allowing you to produce deeply explicit, sensual, and fully uncensored video generations with fluid motion dynamics, authentic skin texture, and rock-solid character stability across frames.
Updated T2V Generation - V2
Major improvements to Text-to-Video handling. It now clearly and reliably generates Pussy and anal details directly from text prompts without breaking composition.
Before anyone drops the inevitable "prompt please?" in the comments,no I’m not typing out prompts for dozens of clips. I'm way too lazy for that kind of manual labor when I'm generating hundreds of files at a time. That’s why I linked the prompt skill right there. Just download it, let the engine handle the formatting, and use a little bit of your own imagination.
⚡ What This LoRA Unlocks
True Anatomical Coherence: Natural breast physics, explicit nipples, realistic body contours, and tactile skin-on-skin contact without the typical melting or rubbery AI artifacts.
Complex Erotic Motion: Smooth physical interactions—from delicate disrobing and slow caressing to passionate kissing, heavy breathing, and intimate couple dynamics.
Dual-Mode Compatibility: Fully tested and calibrated for both Image-to-Video (I2V) and Text-to-Video (T2VA).
🛠️ Training Configuration
Base Model: MiniMax Hailuo H3 (Video)
Training Architecture: LoRA
Hardware: H100
⚙️ Recommended Generation Settings
To get the cleanest motion without artifacts or warping, stick to these baseline parameters:
Recommended LoRA Weight:
0.3 – 1(Start at0.5for the sweet spot between motion responsiveness)Sampler: Euler_ancestral
Sampling Steps:
4-8with (turbo LoRa)CFG / Guidance Scale:
1
🔑 Trigger Words & Keywords
Incorporate these triggers into your subject descriptions and action progressions:
Core Triggers:
nude,topless,bare breasts,naked,completely unclothedAnatomical Tags:
perky nipples,areolas,exposed cleavage,bare buttocks,toned abdomen,subtle skin sheenTactile & Motion Keywords:
caressing bare breasts,intimate kissing,slipping off robe,disrobing,breathless sigh,skin-on-skin contact,arching back
🧠 MiniMax Prompting Skill
We built a dedicated AI Prompt Director Skill that turns your AI assistant (Antigravity, Cursor, Claude Code, ChatGPT, or Claude Projects) into an automated video director. It automatically structures the official 6-section Full-Reference prompts, locks millisecond lip-sync tags and eliminates every common prompting bug before you hit render.
👉 Get The MiniMax Prompting Skill For Free [LINK1]
👉 Get The MiniMax Prompting Skill For Free [LINK2]
⚡ ComfyUI Video Workflow (Step-by-Step Civitai Guide)
Not sure how to wire the nodes, load the LoRA, or pair the Euler_ancestral sampler with 4-8 turbo steps?
We published an in-depth tutorial article right here on Civitai breaking down the entire setup from scratch: optimal VRAM memory allocation, starting frame generation with Flux, turbo LoRA step pairing, and node connections.
👉 Read The Free Step-by-Step ComfyUI Workflow Guide on Civitai Articles
👉 Youtube Tutorial (Proper Tutorial will be updated soon)
👉 Workflow Link
🚀 Join The Midnight Lab (Latest AI Updates & High-End Workflows)
Want access to our complete, ready-to-run production studio?
When you join The Midnight Lab, you skip hours of node troubleshooting, missing custom dependencies, and broken VRAM bottlenecks. Members get instant access to:
The Plug-and-Play MiniMax H3 ComfyUI Production Workflow (fully wired with LoRA loaders, Flux starting frame engine, and upscalers).
Weekly AI Updates & Exclusive Drops: Stay ahead with our newest private LoRA models, advanced workflows (The Muse Engine, 4K Influencer Studio), and cutting-edge video techniques before they go public.
👉 Join The Midnight Lab on Gumroad & Access The Full Vault
Description
Updated T2V Generation: Major improvements to Text-to-Video handling. It now clearly and reliably generates Pussy and anal details directly from text prompts without breaking composition.
FAQ
Comments (34)
So this new V2 is the real V2? right?
yes
@The_Midnight_Lab Love it! .53 is the best setting for me. Thank you for all that you did to put this together. Will be tipping you
Hell Yeah !! Now we're talking abt !! ... and it's not 88k + Buzz to get it ! Love yah !!!
Hello again. Thanks for the update💕, but the situation regarding the T2V has worsened😲. The new audio is excellent, but the video quality has dropped. Furthermore, v2 is incompatible with many other nsfw loras. Until v3 arrives, i will only use it for i2v.
I didn't test with other loras, for me it worked well from 0.5 to 0.7
And yes minimax performance best with i2v and r2v
'AI Prompt Director Skill that turns your AI assistant (Antigravity, Cursor, Claude Code, ChatGPT, or Claude Projects)' anyone got any idea how to bypass the nsfw restrictions those ai assistants have?
In my limited testing, Claude Sonnet / Opus are not prudish at all. i regularly use them for very nsfw projects
try using local agent
V2 seems to be overfit on sex movement, the subject looks like riding an invisible penis even if prompt doesn't mention it at all. There are no such problems with V1.1, which I personally consider the current best NSFW lora. I especially love that it can be used at 1.0 strength with R2V without causing any artifacts: no quality drop, no identity drift, prompt following stays good. Though I prefer the hybrid 15-block version, the vanilla R2V also works well.
It's true, version 1.1 runs more smoothly for me than version 2
I also noticed that V2 is less ref-friendly than v1.1 In particular it adds anatomy artifacts (phantom fingers etc.) and changes body proportions, at least at 1.0 strength. But there are no such issues with v1.1
I haven't tried v2 with fl2v as I almost never use it. Maybe it works better with T2V/I2V workflows.
This is the best loral 💪 Are you planning to make a version for Qwen-Image 2.1 Edit/Text?
Yes
I'm getting much better results with V1.1 compared to V2.
I noticed the clitoris is always positioned on the top, whether the character is facing forward or backward. Could this be optimized in the next version?
That's the new Clitoris v2.0. It's the same orientation regardless of the direction of the woman. It's for men that just cannot perform cunnilingus properly. You're welcome, Ladies.
For those who have used it, could you share your experience? How does it compare to Redcraft, which seems to have disappeared sometime ago? Are there any specific limitations on the models it displays?
haven't used v2 yet, but the earlier versions have been great. Sometimes it's perfect on its own, other times combining with other loras like Mystic XXX v4 is needed.
fwiw I'm pumped to use v2
I have never had luck with any version of Mystic (which I think goes back to SDXL or at least Flux 1). That model seems to only be useful for people interested in focusing on females. I have had better results using 1.1 of this lora than with Mystic and it does seem to improve the results over no lora at all. However, male genitals quality/consistency is still not good. I have tried some penis loras and Naughty Times (the 512 rank version) but male genitals still seems to be a strong area of weakness. Hopefully V2 of this lora will be better for that. Another thing I am trying to overcome is the model's desire to give all guys small waists and butts, even using a starting image. It's harder to make a more burly (not fat but bulky) man. Normally that's not my taste but it would be nice to have the option for variety.
I also got the dreaded "hole in the testicles" issue with 1.1 of this lora yesterday again. Mini struggles with the concept of an anus. With tediously lengthy prompting I was able to move it upward but it was still too low.
If combined with any other lora, it just makes a blurry mess
I haven't tested V2 but I have tested 1.1 a fair amount and I haven't had that issue. Things you might try: HeunPP2 and SA Solver Pece samplers, both using Simple scheduler. No turbo/accelerators. 44–46 steps with HeunPP2 or 55 with SA Solver Pece. Don't use INT8 convrot nor FP8 base model. Use the full 16-bit version. The issue I have been having, with all of the loras so far, is poor male genitals quality/consistency. It isn't at the level of Wan 2.2 at its best at all. However, I haven't put in as much time testing as with that model. I really think Mini should have MoE like Wan 2.2 does. Trying to cram everything into a single model is a mistake, given the super-low VRAM limits we have with consumer cards.
I literally learned how to buy yellow buzz for this LoRA and it's worth it. It fixes so many issues with NSFW videos turning into body horror when using the base model.
this work for ref2v?
no, blurry outputs
The clitoris is inverted)
meh, I think it's pretty much established nothing beats Eros for Minimax in terms of making porn
This the best general NSFW LoRA so far with MysticXXX's one. However it does add a man's voice even if prompted for mute/silent. Hope we will have a silent one for V3 ♥♥♥
incompatible with spectrum makes body horror with that sampler conflict, any fix?
I plugged it into an extender workflow: https://github.com/tritant/ComfyUI_MiniMax_H3_Extender
Tested with strengths of 0.5, 0.8, 1.
Used modes:
minimax_h3_ref2va_pruned_int8_convrot.safetensors minimax_h3_fl2va_pruned_int8_convrot.safetensors
BJ action looks fine (the cock looks ok ish, and the mouth movement is fine) but pussy and ass look quite a lot worse than the previews.
Is the workflow/models I use not right for this? Or is it just a skill issue (bad prompting)