Early / Experimental Release
This LoRA was trained for MiniMax H3 Ref2VA to reproduce BLACKED-style adults films, positioning, camera behavior, and overall visual presentation.
The dataset includes a wide variety of environments, angles, lighting conditions, and compositions, so the model can sometimes produce very cinematic-looking results when paired with the right reference image and the correct prompt.
Current Limitation
The main limitation of this version is fine genital anatomy.
After multiple training runs with progressively larger datasets, motion and overall visual quality improved significantly, but small anatomical details can still be inconsistent depending on the seed, reference image, camera angle, and LoRA strength.
In particular, anal anatomy may sometimes appear simplified or partially missing.
Temporary Workaround
During testing, I found that dedicated anatomy fixer LoRAs, Models like, Synth Pussy version 0.5 or FineLoRas - Pussy & Anus (preview) can noticeably improve this issue.
Combining this model with anus / genital anatomy fixers often restores missing detail and produces a more complete result.
This is only a temporary workaround. My goal is for future versions to improve anatomical consistency without depending on additional fixer LoRAs.
Future Development
The current dataset is around 250 clips, and I plan to continue expanding it toward ~600 carefully selected and captioned clips if possible.
Future training will focus on:
Better anatomical consistency
More camera-angle variety
Better fine-detail preservation
Improved generalization
Maintaining the cinematic quality of the current version.
Support Future Training ❤️
Training and developing AI models requires a significant amount of time, compute, storage, and often paid external GPU services. If you enjoy my work, any support through my Ko-fi is greatly appreciated and helps fund future training, larger datasets, and new model releases.
☕ Ko-fi: here
Description
FAQ
Comments (11)
Great work!! Did you train only for the positions seen in the samples (cowgirl and riding), or are there more positions in the training data? It would be great if you added more positions for the next version, such as (blacked prone-bone, blacked full nelson, blacked double-penetration, blacked blowjob).
Thanks! Yeah, I definitely plan to do more poses. I already have separate folders/datasets for different positions. Training multiple poses into one single LoRA can get pretty tedious, so I prefer keeping each one more focused and release them gradually
@Maviri That's great to know! You're right. ;)
Your training notes re: detail hit home for me. I have the same challenge, big time. I find adding high resolution photos alongside the videos can help. I also use "detail buckets" of 17 frames to try to get it to learn the right shapes, but it's not good at latching onto smaller details at all. I think it has to do with the fact that the model is distilled and training further compresses things.
If you have very high res videos, consider cropping dupes zoomed in at lower repeats, that also may help.
Thanks for commenting and for the recommendations! I really appreciate you sharing your experience, ❤️👍
And yeah, I’ve been testing that too. A part of my dataset already consists of close-up shots focused specifically on the genital area while the action is happening, so the model is definitely seeing that detail during training.
Even with that, I’m still running into the same issue. In some generations it looks like the anatomy is starting to form correctly, but the final result ends up too smooth or under-detailed, almost like the model knows where the structure should be but doesn’t fully resolve it.
@Maviri so I played around with anatomy for a really long time, one of the reasons I didn't release much for a 1-2 week period, it wasn't even to release but to learn... I unlocked detail in a set with LOTS of high res images at max training resolution, but the way MH3 trains is odd, I'd get one epoch out of a ten showing things mostly right, with the adjacent epochs much worse in weird ways. It felt like the window of proper details was tiny. Lowering the LR didn't help, it actually seemed to make things worse, it would still hit the target and then overshoot or undershoot seemingly at random.
I largely am not even attempting to get those details perfect anymore, I can't train the same LORA 2-4 times and spend hours searching for perfect details when the primary concept works well across the board...
My only other note is that I find MH3 needs 7e-5+... a few identity LORAs I trained needed 1e-4, which I know is the default, but I've always trained at much lower LRs...
Good luck and hit me back if you find any other tips!
how To install missing nodes??
with ComfyUI Manager
@Mavirinot found
Would love extreme sloppy bj / deepthroat as next concept. Long strands of mucous-y bubbly saliva probably hard to reliably get prompted.
Gah DAMN....Thank you!