This is a lora test for split screen for groping. Trained with a very small dataset so it is very limited in freedom. This is V1. Might work on a V2 if people think it is worth the time.
It should work for both t2v and i2v
I2V Prompt example:
integrated_multimodal_description: [Shot 1] The video presents a split-screen view of a single person. On the left, a young woman with long dark hair and bangs is visible from the chest up, wearing a white collared shirt and a blue and white striped tie, with her eyes closing at every touch for pleasure and her head tilting up and down in synchronization with the moaning expression of her moving mouth. On the right side of the screen, a close-up shows a man's hand actively lifting the woman's skirt and then inserting his hand directly inside the woman's underwear for gripping, rubbing, and sliding up and down against her pussy iside her panty with continuous, deliberate physical movement, tou_ch, creating a dynamic and engaging action sequence. The lighting on the left side is brighter, while the right side is and focused intensely on the movement. The camera is dynamic on the overall split-screen composition and the left and right screens are related in movements.
overall_soundscape: The audio consists of the rhythmic clatter of a train moving on tracks, accompanied by a low, continuous hum, and the woman voice in pleasure.
non_diegetic_music: na
Description
Small dataset. Rough test
FAQ
Comments (30)
There are three files; are they all necessary?
No, it's just different training step safetensor file.
@Pat3dx Which is the one you have tested?
@AI_man2025 not tested yet.
But, just use the full trained safetensort file:
minimax_h3_split_screenx_rev1.safetensors
Thanks for sharing! I have a question: which model should I download? I see that it's version 3.is this intentional, that the hand is a male hand instead of a female hand?
It's meant to be a male hand I think, but that doesn't make sense given there's no sign of a man's arm in the other split screen
@AppleOrange I dont think. The woman on the left panel clearly masturbate. So logicaly, it must be the woman hand in the right panel :p
The lora correctly keep the outfit consitent for the hand, but not the genre :p
But it should be easy to fix, just lower the strengh of the lora to 0.8-0.5, and explicitly add female hand, or "she masturbate with her right hand".
This should fix the lora issue.
One thing that also trigger video model to genrate female hand, is to add, she masturbate with her right hand, she has red nail polish.
This work perfectly with Wan22, LTX, to always trigger female hand in the video.
I often use this tips when i'm genrating futanari video. where the futa need to jerk off her own penis or the other girl penis. And it work perfectly.
I noticed this too but I just thought, maybe she works out? Plus we can't see what's beneath the persons panties so..? Crying Game sorta deal.
yes, intentional. it would be the groping on asian train style. But I mean H3 does well even without the lora. I thought maybe some can notice a difference and make use of it. I can still remove the post if useless. In theory you could prompt more hands on both split scenes. you gys have to look at the asian style groping videos
@nnewtoai719 ok thank you
At 2.31GB is a bit unexpected.
Ran locally? Because of this below (using the official model). But it somewhat works... doing some other tests.
[ERROR] ERROR lora diffusion_model.blocks.40.adaln_proj.linear.weight shape '[96768, 2688]' is invalid for input of size 774144
Good test, but I think a well trained lora would handle split screen on it's own. More versatile groping would probably have more appeal.
Wouldn't it be more convenient to merge everything using editing software in post-production—much like the workflow for the TV series 24—rather than the current trend in the community of trying to cram every single step into ComfyUI in hopes of a one-shot generation? Some people seem to want to move the entire Adobe software suite into ComfyUI.
Iterating upon this could be interesting. Like one for ass grab as well.
'grab dat ass' prompt works for me
@nidude Could you upload an example?
@Rosettasees my wife wouldnt appreciate that.
Stacked with a another lora or two, this is very versatile with REF2V. 👍 TY
Actually I found the best two loras for this. Anyway. GLHFDD
@nidude do share, and @Korros can you use this in a non spilt screen way?
@nidude cab you share the use you do? I am so curious :-)
The lora is 10x bigger than necessary
bro why 2.3 GB already prices are high on for SSD and RAM
Without it the prompt works very well. I really cannot figure the need for it.
Can you add multiple hands without the lora?
You can suggest maybe what the main base model can not do so that I can fine tune the lora in a version 2?
I would rather accept constructive feedback if possible?
Same testing as usual, Local Gen on a RTX 5xxx 16GB.
I ran your example prompt and LORA with stock H3 FL2VA and REF2VA, with or without Turbo and I'm getting the LORA Error I documented earlier and a result similar as in the gallery. Then without your LORA, your prompt generated very good results similar to your sample.
Also the 2.5GB is rather weird, generally action LORAs tend to be sub 400MB as per the LORAs I accumulated.
I'm using the ComfyUI (latest) and their Templates.
If I had a hint of your template/setup I could have tried it... but this wasn't wasted effort and I have yet to try training a H3 LORAs myself so I cannot provide better guidance pass this. Sorry.
I don't know about the range of freedom this has.. but ideas like this are very interesting
Definitely want to see more ideas-guiding loras like this
Thank you
I didn't use a big enough dataset. Also H3 seems can do split video. IT is hard to understand where H3 lacks in such videos and where we need to fine tune the min model with the lora. If you can give me a feedback of what needs to be fine tumed, maybe we can make this lora some how better or different , etc
One of my many challenges, is understanding where H3 can do things alone and whre we need a lora to fine tune it