MM H3 reference video ref2va lora
MM H3 lora trained on the reference to video model (ref2va), fun little project for me so I could make stupid videos. I'm using euler/beta. I would be cautious about mixing some other loras (test first) because when I try them they make videos look poor quality or cause smearing. I would also test without turbo first (I dont use them). use only ref2va trained loras. Don't forget with ref2va you can supply short 512x512 reference videos for your scene, that really helps a lot.
FYI: I dont use turbo lora, for various reasons, and I dont test loras with turbo loras enabled. I think it makes the distilled lora training related issue worse, rendering the lora useless.
Grok H3 Prompt Skill: https://grok.com/skill-link/6dc971af0dd0458b12ef523fa62696bf
How does a reference model work?
It doesnt work the way you are used to with traditional models. This is a reference model, it literally means you need to give it reference images because the lora and the model itself is trained on reference images. If you want a nude woman to suddenly run into the scene, give it a reference nude woman. If you want someone to look over and see two people fucking give it a reference image of what fucking looks like. Then let the lora animate it. The lora was trained on "take these reference images" -> "generate a video using them". it was NOT trained on text to video, it cannot generate a cock out of thin air, it needs a reference image to animate it. Otherwise use the fl2va model because it can be trained on pure captions. So if im doing a handjob scene, I might give it the main photo of the scene, a reference close up cock image, a reference oily tits and reference handjob technique photo. Doesnt mean I need all those, I could get away with one image but I cover my bases so the video turns out. The job of this lora (and what it was trained on) is to take images and animate it, it was not trained to create anatomy that doesnt exist in the references. It can sort of do it but it will look incorrect (deformed cock head or something because the angle of the cock is different than what it may have seen).
I put in the comments for each of my v1.2 videos how I created it. I uploaded a "what does a nude woman look like from all angles" photo to the gallery, reference it in your generations so the model doesnt have to invent anatomy as the camera is moving around. https://civarchive.com/images/140155060
you can even give this model video examples of how to fuck, or audio references of what it sounds like, the sky is the limit if you are creative. But its a bit of work compared to the simple fl2va model. this lora was trained on how to make reference images move, hopefully that helps.
You need GOOD prompting, use the grok prompt skill below unless you want to write 1000 word essays.
the two versions of this lora are described below, I would try 'sexytime' variant first.
Join our Discord: https://discord.com/invite/GCrr4Cj3D
WARNING: for the AfterMidnight ref2va lora you need to use euler sampler and beta scheduler, or you will get weird crap like audio artifacts or turbo tits. Listen to some of my videos and you can hear odd audio noise or see strange crap, thats from me using res_multistep. dont be a like me, cheers. this video uses the correct workflow https://civarchive.com/images/140029319 :)
Make sure the lora passes to your Basic Scheduler and Guider nodes or nothing will happen
Workflow: download video and open in comfyUI https://civarchive.com/images/140064227
Lora "Flavors"
there are 2 flavors of this lora (same dataset different training styles), one only one at at time:
sexytime: strength 1.0. Was trained for sex scenes and coherent motion, probably closer to what you want.
softer version: strength1.0, wasn't pushed on motion as much, focused more on detail. I run at strength 1.0. Was trained on detail and surreal style, think like crisp outputs and fantasy (flying dicks?)
Model: MM H3 ref2va bf16/int8
Training: mixed bag of 1500 videos of various sex acts, close ups etc
this is my first attempt at a ref2va mm h3 lora. Still working on my training pipeline for this one (I dont use ai-toolkit)
Description
v1.2
FAQ
Comments (30)
Is it possible to make a sex with visible female and male anatomy without First frame? Only character reference and the sex description? I can't get good results this way.
its not a first frame model, its a reference model (read the description on the main page for this lora). you want to provide reference images or videos to the model, thats how ref2va works. if you have a character reference that is fine, include several shots of that character and reference them in the prompt using the prompt guide
@sexgod1979 I came back the next day and read your description; it was very helpful. But I have to be honest: even with various reference images of sexual positions and genitalia, and using GROK for the prompt, I can't get the character's genitals to look right. If I need an explicit image of the genitals already positioned on the character, the Ref2VA mode loses its point a bit; plus, in my tests, toggling the LoRA on or off made almost imperceptible differences (like a missing clothing accessory or a slight change in camera angle).
@highlandertheonde397 whats your setup like, using a turbo (dont)? give me some background or hit me on chat and I can problem assist.
dont know if you use a turbo but
FYI: I dont use turbo lora, for various reasons, and I dont test loras with turbo loras enabled. I think it makes the distilled lora training related issue worse, rendering the lora useless.
@sexgod1979 Thanks for your willingness to help. Here’s the situation: I’m using the default ComfyUI workflow without Turbo LoRA. I’m connecting the "Load Diffusion Model" node directly to the "Power LoRA Load (RGB Three)" node, followed by "Sage Attention" and "EasyCache." 25 steps. 10 seconds of video. My goal: to use image references to create an explicit video of a character having missionary-style sex with a generic man who is partially visible on the right side of the frame. References:
- Picture 1: Character sheet/multiview;
- Picture 2: Close-up of the character's face;
- Picture 3: Reference for a nude female body (torso);
- Picture 4: Reference close-up of a vagina being penetrated at an angle very similar to the intended video shot;
- Picture 5: Specific tattoo located on the leg (hidden beneath the clothing in Picture 1);
- Picture 6: An image of missionary-style sex (matching the intended video);
- Picture 7: Close-up of a generic penis.
I’ve tried countless prompts on Grok and made manual adjustments myself, yet the genitalia still look like the base model's monstrosity—with only a few slightly better results depending on the seed. Overall, though, the penetration angle is wrong, and the anatomy is poor.
@sexgod1979 Okay, update: I tried the following—I ditched my complex prompt involving specific movements and camera angles and simplified it, essentially just telling it to take the character from <Picture 1> and put her in the pose from <Picture 6> (a missionary sex shot), while describing the movement as fast and firm. The result: The model simply swapped the head from <Picture 6> with my character's head, so the pose, the vagina, and everything else are identical to the sexual reference material. At that point, I expected the LoRA to animate correctly since everything was in place, but it didn't. It basically does nothing; I removed the custom nodes, used only the standard "Load LoRA" node, switched the Basic Scheduler and KSampler settings to the ComfyUI defaults, and still got nothing. Using the same seed, only micro-movements differ when using the LoRA, but the animation of the penis entering remains terrible—looking as if it's being thrust into a gelatin pillow—and the vulva is interpreted merely as a soft hole, just like in the base model. Could you try replicating something similar with visible anatomy and explain exactly what your workflow looks like? Because right now, it seems like using the LoRA makes no difference at all.
@highlandertheonde397 thanks for this, I am looking into this. I suspect I know why, but I have to test it and possible adjust something in the lora to address it. ill circle back on this
@sexgod1979 I know it’s not your LoRA, but perhaps this helps with the solution. I achieved my goal masterfully using the [MMH3] Mystic XXX V2 LoRA at 0.75 strength; it isn't specific to Ref2Va, but by using reference images for the sexual act and genital close-ups, the result was perfect. If that is your aim, I hope you manage to create something sublime with this LoRA. Thank you for your work.
I'll take all three, good sir.
SexGod "After Midnight" MM H3 ref2va = 2D-Holodeck
Is it just me, or do MiniMax LoRAs never work as expected? I see no difference when adding them
Just you, seems like people have tough time wiring loras with the correct model I guess. YOu can join our discord if youre getting stuck, we lots of example videos
I got to agree with you, but only for some extent. Prompting is what matters, normally I will spend like 1 whole day just to revise and test out 1 single scene until I got the result that I want.
@tickaestore799 one issue is that this is a distilled model, unlike LTX 2.3 we were not given the weights of the undistilled model to train with, so it has to fight the base model often times the base model will push through. normally we would want the undistilled model to train against, thats why there are not 'distil loras' for this model because it already is. You end up trying to train a lora against something that has already strong guidance. that could be more what the OP is getting at, but it is what it is with this model and we have to live with it.
there's a common workflow available on civit ai that is broken and never wires the lora loader to the chain. check your workflow.
I work at wan2gp and when I downloaded SexGod LTX23 DoggyStyle earlier v2_5 I was absolutely delighted. It was generally the best thing I downloaded on this site - SexGod LTX23 DoggyStyle accurately and obediently worked with all your fantasies, even very far from DoggyStyle style. Today I downloaded SexGod "After Midnight" MM H3 ref2va and... it doesn't work at all. No visible result. Sadness.
I havent used wan2gp so I dont know, im training on the comfyUI base model. It works fine for me, but its hard to know how everyone setup, workflow, model different as well as their nodes etc. I can only confirm that my workflow, model and setup works as expected. I also have a lot of experience with these models since im training them and know how to get them to do what I want them to do. We just tested this with some users in discord using example ref videos and are getting good results with it. Unfortunately really hard for me to dig into all the individual setups. imusing comfyUI, my workflow, unpruned int8 model and the lora at strength 1.0 to get the videos in the gallery.
also the model itself is distilled that could be a factor, since we are training on distilled base models, but not having any experience with wan2gp it would be hard to comment on it as I only use comfyUI
we do have several people using wan2g with the pinkcherry beta-0.6 instead of the lora but that is fl2va model and it requires downloading the checkpoint and setting up a json file in wan2gp, so a bit more work than comfyUI
@sexgod1979 In any case, I want to wish you success. Amidst all the rubbish that floods in here, some of your works represented a real breakthrough.
@ilya808rolf994 thanks, I think all of us are still trying to figure out this model since the undistilled weights were not made available to us, so we are working to "un-distil" this model to train it better. great fun but not the easiest task
Tried it on Wangp v12.53 with Ref2VA Prunned 20B and it worked, audio is not the greatest. I tried Aftermidnight lora strength at 0.5 and at 1 and this turbo strength to 1. (https://civitai.red/models/2837571/minimax-h3-turbo-loras?modelVersionId=3219346). Ran it on 6 steps. Ran at 540P. Used reference image for clothed and another reference image for unclothed.
On the html/web inter face for Wan2gp I did have to go into the tab Configuration then subtab Performance and set it to a lower system Ram profile than I actually have for some reason. I didn't have to do that to get Wan2.2 or LTX2 to work.
After reading the description, this is literally just i2V but with infinitely more steps, processing time, and quality degredation. What is even the point of a reference model if it can't actually infer anything from reference?
yeah now that may just because my lora trained on image conditioning, its not to say I or something else cannot fine tune this model. I think likely thats the way I will go with this one is to fine tune it rather than rely on a lora.
I'm looking for anything for the reference model myself. The flfv model is handicapped massively for my uses and has absolutely no appeal to me. The ability to have Character1 from <Picture 1> be in a different location with Character2 from <Picture 2> in a pose resembling the person from <Picture 3>, and have that all work along with extensive details about what all they are doing with anonymous references like 'Character1 does this and Character2 does that' after invoking them in a similar way and all of it works perfectly like magic first try; absolutely amazing. After using multiple character references effectively, start and end frames seem like the stone age.
So any Lora I can get for that is good to me personally
i've tested all of ur lora for minimax and i found that your lora unlike other loras, it's only work by its own and like you said it will ruin the output if we even add a single other lora. Is there a chance where u gonna update it to something more friendly to other loras?
This one was trained on the ref2va model weights, so its possible the other loras are trained on the fl2va model. I noticed when I use other loras on the ref2va model even without mine I get blurry outputs. Thats my guess. If youre doing fl2va model you can use the pinkcherry 0.6-beta and use other loras as they work
Is there a list of the top ten settings(location training videos taken at) so we can prompt them in the videos? I know it is impossible to list 1500 locations. Feel like those locations would look more clear in the videos.
There used to be an awesome plugin for prompt generation I'd us for a1111/forge, that had a great list of all these common nouns and verbs and stuff. I would guarantee there are similar nodes for comfyui.
I can't find the one I used, but the idea is the same: https://github.com/tritant/sd-webui-creaprompt/tree/main/csv,
or specifically: https://github.com/tritant/sd-webui-creaprompt/blob/main/csv/3_3Places.csv
Find any type of repo like this and they'll have all these CSVs with lists of common words. You could also look specifically by your CLIP and see if people have extracted or provided the words that it should know
This works very badly