New ref2va file to download (experimental, no audio trained)

This works for Ref2VA! Use a hybrid model! AND IT IS FANTASTIC!

@DaddyWolfgang Ive also noticed this that many minimax loras trained on fl2va also work on ref2va good

@DaddyWolfgang what do you mean by 'hybrid model'? Just using this I2V lora in a ref2v workflow?

Models tested:

Model: minimax_h3_ref2va_pruned_int8_convrot

ref2va: Cannot be used

MysticXXX_MMH3-V4: Cannot be used

MysticXXX_MMH3-V3: Cannot be used

MysticXXX_MMH3-V2: Works correctly

MysticXXX_MMH3-V1: Cannot be used

Hybrid model: minimax_h3_hybrid_fl2va_ref2va_b20-49-int8

MysticXXX_MMH3-V4: Works, but character consistency is lost.

MysticXXX_MMH3-V3: Works, but character consistency is lost.

V2 works normally and maintains the character consistency I need. However, V3 and V4 either cannot be used with the regular ref2va model, or they lose character consistency when using the hybrid model.

Could you share the daswai nsfw model? It got taken down

could u share the results?

thanks for the details

Models tested:

Model: minimax_h3_ref2va_pruned_int8_convrot

ref2va: Cannot be used

MysticXXX_MMH3-V4: Cannot be used

MysticXXX_MMH3-V3: Cannot be used

MysticXXX_MMH3-V2: Works correctly

MysticXXX_MMH3-V1: Cannot be used

Hybrid model: minimax_h3_hybrid_fl2va_ref2va_b20-49-int8

MysticXXX_MMH3-V4: Works, but character consistency is lost.

MysticXXX_MMH3-V3: Works, but character consistency is lost.

V2 works normally and maintains the character consistency I need. However, V3 and V4 either cannot be used with the regular ref2va model, or they lose character consistency when using the hybrid model.

@alcaitiff  Nah :D this is over the place I told already in my comment. We need physics and smooth movement loras nothing else for rev2va

For some things, at the right strengths, they can help. But even T2V or FL2 I can get anatomy that doesn't start on screen to exist (like anuses and glans and proper vaginas) with just Mystic v3 or v4 alone. But sometimes it might not be as perfect as you might want, and certain loras have improved some of my outputs. However, most other loras clash with the audio, so even if the picture improves, it still results in most sexual noises reverting back to those crunchy noises.

One hack -> In reference images attach both penis and vagina seperatly.

The easiest way is to reduce the mystic strength and try adding another lora at a low strength that has more control over it. Using Krea2 or F2K to create starting/ending images is the most direct route.

A lot of us focus on the ref model, but depending on the concept it requires separate work or tweaked setting. We just have to be patient.

Also there are some known issues with the reference model that is on the radar for MiniMax to fix.

You mean the one that's more complicated and requires just doubling up on the training? I feel like people don't know what they're asking sometimes. You can see that v4 is quite close to finished, and yet you complain there isn't a whole second model. There are full blown custom finetunes for the ref space, including Eros and PinkCherry. It will probably exist... but let the dude finish.

Ref workflows work with FL2VA model so...

this kinda work well with the Ref models if you tweak the setting right

@JJoker375 And what's your advise on tweaking them?

All my V4 are working as expected. It is a little stronger than v3, it needs to be 10% lower in T2V (FL2V have only tried a few at max strength and worked perfectly). For T2V though I use it at 0.6-0.8, the turbo lora at 0.8-0.9, res_multistep, 12 steps. It might have too many steps or the wrong strengths for whatever turbo lora you use, though. Im using some older 4step-850step-ema lora -- I did try newer ones but never got as good as results. I have tried lower step counts and different weights pretty extensively... but I could definitely just be in some weird valley in the sampler regression.

v2 is the best

@makiaeveli im not using Turbo LoRA at all , on version 2 stereght 1 and 20 step working fine but on v4 is horrible the total anatomy is messed up with steght 1 and res_multi step simple

@agapes yes I can reproduce the exact same things at strength 1. Like total blobs of human flesh. But simply turning it down to 0.9 or 0.8, it immediately snaps into reality. each prompt I seem to have to find a sweet spot. Some are quite low, like 0.4 or less. It still keeps anatomy and movement. Also at 20 steps you could try a turbo lora at really low strength (like 0.2-0.4) and it will likely help with clarity and movement.

@chaiwala008215 I also agree v2 is really good I tried v3 and it broke my prompts lol im afraid to try v4 if i am being honest 😂

Does for me at str .4

Sounds like default model. I got once nice slurping sound, but forgot what I wrote in prompt

"low quiet wet sexual sucking" or some variation to lower the loudness and harshness of the crunch. Once you lower the volume via prompt it stops sounding like food (more of the time, sometimes it'll still do it, and some longer prompts will bleed the sound back in).

same for me

My god man . I was going crazy because I was getting shit results with v3. I thought I accidently changed something in the workflow.

Can you provide a sample prompt and workflow that you use? I cannot get ref2va work with this lora. Thx.

yes, at low str it gradually turns into a vulva

ye ye

100%

Yeah, whatever you prompt, or use as reference, their titties get bigger.

they are sadly really too big. It's impossible to counteract this in the prompt. so you have reduce the weight to 0.4

same, blowjob looks weird

Me too. But not tongue, it's something fluid I think.

Were you able to fix this?

@alcaitiff Likewise brother.
We should consider working together for the benefit of the community 👀

@HearmemanAI @alcaitiff Both our goats. +1 on this idea haha

download someone else's post/video, the video itself has embedded config and everything else you need. diff or reproduce and then try again.

Try lowering the influence srt of this Lore. In some scenarios, if I set the str to 1, the quality drops sharply.

Yeah, I've seen the exact same thing. Spent the day testing. Using ref2va model and WF so quite a bit of setup and movement in the prompt. V1 nails it.

Same. V2 is better for me.

got okay result with v4(0.5-0.6 strength). lightx2v 8 steps lora. 8-12 steps. multires/euler with simple.

grok is surprisingly competent. downside is the usage limit. other than that i just use prompts i've used before and tweak the template

Or you can use local Gemma 4 12B heretic to make nsfw prompts

@marmoth5 i cant afford to pay for another model, claude is already planning to leave me homeless lol, thanks though i might check it out in desperation

@Stefan_Falkok installing now via hermes, thanks 

did you feed Claude the instructions from the Krea repository? there is a code to activate the skill.

@jamesianv253  let me check and report back

@Stefan_Falkok do you know another models like Gemma 4 12B heretic to make nsfw promts?

@mesugaki_csk new Qwen 27b models maybe

@stealthy_007 if you prefer local, i use gemma 4 26 A4B uncensored heretic. you feed it the prompting rules from hugging face and it does a pretty good job. I do this when my grok decides to stop answering

@marmoth5 Hi there, well that's awesome. Can you share wf or give more details where/how you get that to work? are you feeding a starting image and asking for a prompt? How would that work in comfyUI*?

Here it is: GitHub - pytraveler/MiniMax-H3-Prompt-Rewriter-ComfyUI: Personal adaptation of the lightx2v/MiniMax-H3-Prompt-Rewriter-LoRA project for ComfyUI · GitHub
Couldn't find a better one. You can also use the hint template from here: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md by writing it as an instruction for a suitable LLM, for example, the same qwen 3.5 or 3.8 or whatever

@Delavestra steps:
- download lm studio
- download the model inside lm studio

- paste the hugging face prompting guide (all the text) into the template preset of your newly downloaded model

- now when you talk to the model, it knows all of the prompting structure.

- recommendation : talk to it like you would a chat ai. give it some guidelines beforehand like tell it to do 'temporal prompting'.

if all this is still confusing feel free to dm

Grok 4.5 is cheap.

So I found the best way to solve this.

Thanks to everyone who said Grok, it quiet literally is the most underrated AI. I bought credits and gave key to claude and delegated prompting completely to grok. It costed me $5 and that's good for almost 2 months. Claude handles everything else but for prompting it uses grok only.

Can you tell me the model you are using and the strength of the lora?

@kkmw15 MysticXXX_MMH3-V4-ref2va：0.65

Most of the sample videos here have a workflow attached. Just simply drag the .mp4 into ComyUI

@maierjuergen1976468 谢谢，但我要的不是提示词，而是完整的工作流文件

when you scroll down here you will see a lot of videos. When you see the Option "Remix" here it means that the .mp4 has a Workflow included. Download the video and drag and drop it into your ComfyUI. You will get the full workflow!

@maierjuergen1976468 Thank you very much, but there are very few ones with workflows.

Depends on your GPU. If you're forcing it to offload because it's full of naughty LoRAs, of course it's going to slow down. Show your console to an LLM. It's faster than asking here. And it won't make fun of you. Unless you ask it to.

@Ponder_Stibbons the thing is I only ever use this lora. I avoid stacking loras because I know they can mess things up in many ways. So if this one lora CAN cause massive slow downs then I guess that answers ny question.

@123sirako123521 sounds like you dont have enough vram or ram. Only your wallet can solve that one.

The lora likely pushed your VRAM over the edge and it spilled over into shared memory leading to the slowdown.

Yep. I've experimented with all combinations. Ref model and a few of the hybrids, with 4step ref, 4step flf, 8 step flf. Best results for me so far have been with 4step LoRA at 0.8, 4-6 steps euler/simple. 8 steps is NG imo, but much depends on your sampler choice. But ultimately distilled is distilled. Don't go crazy trying to find one that's going to perform like an undistilled full model run.

感觉目前大部分lora训练到后期都会面临这个音频问题。
社区应该需要更多的时间去研究这些东西。我会继续耐心等待。

same to me

show reference, that's all.

you might also need another lora that focuses on that. H3 doesn't know how to "draw" a penis out of the box.

what's the point of dick loras then? and t2v models manage

idk if you've figured this out yet but i've had good success using just this one. i've used v2 and v4 with 0.8 strength and i describe it as big erect penis. works great

For me, this issue is generally related to an overly large video_shift value. Of course, it may also be associated with the base model you use.

I have the same, do You know why it is?

v2 is the best

v4 es perfecta en R2VA, PERFECTA

i wouldn't be so harsh, but most of the guy critique is correct, only v2 avoid whose issues at some point, but it's motion is not that great.

Are you trying to say anal? wtf

Why type like this

@jazzyreynard285 Yes, it will show up as **** if typed directly.

@playnproto266 Yes, it will show up as **** if typed directly.

Aye it does struggle with getting anal sex right. Hopefully thats fixed for v5.

@louqwer how? Ass and anal are not censored by the platform.

theres a lora for that i think that you can download and hook it up in comfy along this one

@playnproto266 @louqwer  Apparently when you have the "blur mature content" setting on, it also censors horny words like "sex" and "anal" in comment text for you (even though it says it only affects images and video). People who don't have that setting get the uncensored text.

check out my Loras for penis woes. You can try them out for no lora fee with the online generator

Starting by sharing more is already laying the groundwork. Once the momentum is built, good advice will follow naturally, without needing to force it. lol

@lighthorsexajz830 Agreed, thanks! I made a mix of mid resolution with more steps. I think mu 5080 has its limits. But was able to do cool stuff still. About 20% worst than what i see as my goal. But getting there. Thanks for the reply!