This HunyuanVideo LoRa can generate four prone bone camera angles.
It was trained on Windows 10 using Musubi Tuner at https://github.com/kohya-ss/musubi-tuner
I recommend using workflow on the screenshot. Pay attention to video generation settings, especially width and height of the videos (it was trained on 480x270px videos).
You can read how I trained this LoRa here:
https://civarchive.com/articles/10588/how-to-train-hunyuanvideo-lora-using-musubi-tuner
Example prompts:
pr0neb0ne side view, a woman is lying in a bed on her stomach. A man is fucking her in the ass. They are having anal sex. He moves his hips up and down in fast rhythmic movements inserting his dick inside her anus and then pulling his dick out in fast thrusting movements.
pr0neb0ne back view, a woman is lying in a bed on her stomach. A man is fucking her in the ass. They are having anal sex. He moves his hips up and down in fast rhythmic movements inserting his dick inside her anus and then pulling his dick out in fast thrusting movements.
A woman is laying with her legs wide spread. Her vagina is visible.
pr0neb0ne closeup, a woman is lying on her stomach. A man is fucking her in the ass. They are having anal sex. He moves his hips up and down in fast rhythmic movements inserting his dick inside her anus and then pulling his dick out in fast thrusting movements.
pr0neb0ne front view, a woman is lying in a bed on her stomach. A black man is fucking her in the ass. They are having anal sex. He moves his hips up and down in fast rhythmic movements inserting his dick inside her anus and then pulling his dick out in fast thrusting movements.
A woman is opening her mouth and looking at viewer.
Description
FAQ
Comments (18)
you have it listed as a checkpoint instead of as a Lora
sorry, fixed
Not bad. Movements are awfully fast and erratic however. I tried a lot of different settings, changing words from "fast" to "slow" and increasing flow state i think helped a bit. but it's not just this lora, i find a lot of Hunyuan lora's tend to have too quick and erratic movement. I appreciate the work though. I hope to one day figure out how to make lora's on windows locally and maybe understand more how it all works.
have the same problem. if i do get it to stop humping at the speed of sound then it just looks slow-mo instead of realtime slow movement. i think the lora needs a few more steps and more diversity with the speed. have them captioned with the speeds etc.
I had the same problem and may i suggest that you throw the frames through GIMM-VFI and interpolate the frames to make it more smooth and then adjust the frame rate accordingly with video combine node. That should fix the issue with it not looking like realtime slow movement.
Also my 2 cent, I find if I generate small videos you get that jerky motion, double the image size it flips to slow and more natural most of the time.
That's the issue currently with all of these two pass workflows.
in a way it's nice, because if you can tone it down that mean you have a great range of movements, and those fast movements can be meme worthy (even if i'm not really a meme guy myself, I can still see the potential :p)
Thanks, I appreciate your efforts.
it is very hit and miss, but when it hit, it land quite well, (and those that miss are usually rather funny)
on a side note, did you used the virtual environment to use that musubi tuner ? I tried all the steps but I think I didn't do the venv, and got tons of error messages (was trying to train a lora as well on windows, but no luck so far)
Yes, I used miniconda
@guy33 Okay I'll have to try (after I try another one that was suggested to me also) thanks for reply
isn't working well with character loras seems a bit heavy
How is the LORA, with diffusion pipe or other? formed, what are the necessary PC capacities?
"PC capacities" are not words that mean anything to people. Please rephrase in your native language so it can be translated
Teacher, like your workflow, can you directly produce videos from pictures?
Hey just a heads up: HunyuanVideoWrapper can load Musubi Tuner LORA like this one natively with no problem. Native Comfy normally requires conversion but you can use https://github.com/facok/ComfyUI-HunyuanVideoMultiLora (which also allows you to select target blocks as well) to load unconverted ones into native too, I made a PR to this effect which was merged!
idk why I can't run with seageattn, the vram pike so fast, cause OOM ( 3090 ) , sdpa is ok
try diffusion pipe whatever you used is garbage