This LoRa was trained to do short movies made from 1 to 3 scenes. Duration of each scene is 2 seconds (48 frames). In first scene you meet a girl, in 2nd scene she is sucking your dick, in 3rd scene you fuck her in pussy while she is laying on her back with her legs wide spread and lifted up in the air. It was trained on 3 special tokens which model associates with scenes: p0s1 (intro), s8k2 (sucking), f4k9 (fucking). Also videos were marked with black stripe at the bottom left corner with red triangles (intro), green circles (sucking), blue squares (fucking) so if you see them in the video - it's not a glitch.
I will update my Article soon with details how I trained it:
https://civarchive.com/articles/10588/how-to-train-hunyuanvideo-lora-using-musubi-tuner
You can generate videos with different duration. For example, for "two scenes" movie you can increase or decrease amount of frames.
This LoRa can do:
1) Three scenes: 6 seconds video (intro + suck + fuck), trained resolution 480x272, 145 frames
2) Two scenes: 4 seconds (intro + suck), trained res 480x272, 97 frames
3) Two scenes: 4 seconds (intro + fuck), trained res 480x272, 97 frames
4) One scene: 2 seconds (intro), trained res: 640x362, 49 frames
5) One scene: 2 seconds (suck), trained res: 640x362, 49 frames
6) One scene: 2 seconds (fuck), trained res: 640x362, 49 frames
Although 4), 5), 6) were trained on 640x362 I still recommend to generate them in 480x272 to speed up generation time. Start from 10 samples and if you like a video - increase samples to 20.
My personal way of using this LoRa is generate four seconds videos with two scenes only (intro and fucking) for these reasons:
1) It is faster (only 97 frames) compared to 145 frames. I get a video in 60-90 secs.
2) Higher rate of success. It correctly generates two scenes almost 70% of time.
6 seconds three scenes videos - I'd say success rate here is about 40-50% and I don't like to wait for it (around 3 mins on RTX 3090)
Example prompts for 6 secs videos:
1)
p0s1: A woman is sitting on a bed.
s8k2: She is sucking man's dick. She is moving her head back and forth swallowing his dick.
f4k9: A woman is laying on her back with legs wide spread, her pussy is exposed. A man is inserting his dick back and forth inside her vagina.
2)
p0s1: A woman is kneeling on a bed.
s8k2: She is sucking man's dick. She is moving her head back and forth swallowing his dick.
f4k9: A woman is laying on her back with legs wide spread, her pussy is exposed. A man is inserting his dick back and forth inside her vagina.
Example prompts for 4 secs videos:
1)
p0s1: A woman is sitting on a bed.
s8k2: She is sucking man's dick. She is moving her head back and forth swallowing his dick.
2)
p0s1: A woman is standing.
s8k2: She is sucking man's dick. She is moving her head back and forth swallowing his dick.
Example prompts for 2 secs videos:
1)
p0s1: A woman is sitting on a bed.
2)
p0s1: A woman is kneeling on a bed.
3)
s8k2: She is sucking man's dick. She is moving her head back and forth swallowing his dick.
4)
f4k9: A woman is laying on her back with legs wide spread, her pussy is exposed. A man is inserting his dick back and forth inside her vagina.
More interesting examples:
1) 4 secs video:
np0s1: A woman named Margareth is standing next to a sport's car. She has braided pigtails and wearing short slip dress, mini camisole dress.
f4k9: Margareth is laying on a back seat of a car with lifted dress on her back with legs wide spread, her pussy is exposed. A man is inserting his dick back and forth inside her vagina.
Margareth is moaning from sex pleasure.
2) 6 secs video:
p0s1: European woman named Mia is standing on a beach wearing yellow bikini.
s8k2: Mia is sucking man's dick. She is moving her head back and forth swallowing his dick.
f4k9: Mia is laying on her back on a sand with legs wide spread, her pussy is exposed. A man is inserting his dick back and forth inside her vagina.
I can't generate example videos with this LoRa because my RTX 3090 died this night while I was training more steps for this LoRa and the ones I have are too NSFW for civitai.
Description
FAQ
Comments (38)
this is extremely interesting, thank you for sharing
Any example videos?
Never mind sorry! I missed your message at the bottom of the description about your GPU dying. The community appreciates it sacrifice.
cannot get different scenes from your examples, everything happens in one scene
Vids or it didn't happen.
if only people could read
It happened, although have a point
Hi. What do you mean the examples are too NSFW for civitai? There are xxx vids all over the Hunyuan LoRA section.
Sorry about your 3090. GPUs ain't cheap.
I am not sure my videos are aligned with CivitAI "Content Policies" at https://civitai.com/safety
We made it boys
should there also be a workflow connected to this?
I have attached a screenshot of workfow
How about a workflow?
@guy33 yeah, I squinted and well, don't have the Nyan node, but really want it (been trying to get my sageattention back on point, but forgot which node that is). Can confirm it works quite well actually for my own workflow, I guess you don't need one, just the prompting. As a favor, can you upload your workflow though just so I can find that one damn elusive node (or just tell me what package its from ).
@saturngfx https://github.com/zer0int/ComfyUI-HunyuanVideo-Nyan
It is not necessary. I am even using rgthree's power lora loader and all default CLIP/Sampling
@makiaevelio543 Right, mine was just a side question. doesn't matter, I actually tested and SA actually slows things down oddly enough. Probably clashing with other tweaks I got going on.
@saturngfx I haven't tried this, but most of the true optimizations regarding sampling seem to depend on deepspeed or triton, both of which are a hard/impossible to run on raw windows. It likely wouldnt be hard to reset using WSL, but haven't done it for comfy or whatever. Probably just easier to go ahead and setup the docker containers lol
@makiaevelio543 yeah, I use linux, but whenever making any workflow to share, I never include that stuff because its a bit of a pain in the ass to start. not too bad, but people tend to go with easy mode.
Seriously surprised at how well this works so far. Well done.
Woah.
F for your 3090, rest in peace
Can confirm. it helped me to add the beginning to the prompt, before the other two lines:
The 4 second video contains two scenes: the first scene an introduction to the girl while she stands; and the second scene having sex with the same girl.
Really cool! While I haven't had much luck getting your exact workflow to work, my usual workflow with this LoRA applied works well enough.
I've seen it happen a lot where the intro scene will flow naturally into the sucking scene without a hard cut and that's actually kind of cool.
Damn, dude lost his 3090 to make us happier
In the future when we can do much longer videos, such loras like this could make entire scenes
Sucks to hear about the GPU chief. Check the thermal paste if you're mechanically inclined enough to separate the heatsink from the board just to get an idea if it was degraded to the point of not working or inhibiting the cooler from working. I've heard of paranoid individuals reapplying paste every year or two to avoid it happening.
I'll maxx out on this LoRA in the future so that 3090's efforts live on. Thanks for sharing.
man your lora is crazy good, such character consistency between scenes is mindblowing. the only downside is that theres too many skinny asian/phillipine girls in dataset it seems..
Works nicely, great job!. I cant wait until these models are powerful and efficient, we will be able to prompt for much longer scenes in the future for sure. Right now its taking me 10-20 minutes to generate any scene over 8 seconds long. Right now its not really good for pumping out a ton of videos, too expensive still.
WOW. More like this would be amazing.
I only have 12gb of vram. is there a workflow that can stitch clips together? I feel like that would be a revolution!
I use Tipard Video Converter to merge clips together. Works great. Can also upscale, enhance, reverse, make 3D, etc. Low resource use. My laptop isn't powerful enough to use AI on board, yet, the clip merges happen within 10 seconds or so, depending on how many you choose to merge. You can merge a lot together.
@dickardjr good idea, thanks dickard!
@olivereads38255 You bet!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.

