Oh babe, what you're doing?! 😮
Just chilling 😏
Training
320 steps on high noise
2 edited videos (40 frames each video)
HLR + ZCD (These acronyms were created solely to confuse you)
Lots of confidence
$5 (✋🏻🤠🤚🏻)
Usage (Low noise)
This LoRA was trained only on High Noise, which means you'll have to use some Low Noise that knows what a penis is (I used the Low Noise from DR34ML4Y v2).
Got an idea for LoRA? Then let's do it! (or at least try 🤠)
In my repository, I created a model called "LoRA Requests", so if you want to request a LoRA, send videos and photos to help build the dataset for a public LoRA that will be posted here, that's the place to go.
Open LoRA Requests on HuggingFace 🤗
Uncensored prompts and examples
You know that Civitai likes to remove some LoRAs, right? For that reason, I won't post the "Cut to action" prompt. To see all my LoRAs, visit my repository on HuggingFace.
Or visit this LoRA on HuggingFace 🤗
I2V:
Prompt (ongoing, he strokes):
Holding his hard penis with pinched fingers, he moves his hand up and down repeatedly and slowly, stroking his hard penis.Prompt (ongoing, she strokes):
She leans forward and grabs his hard penis in her closed fist. Holding his hard penis with closed fist, she moves her hand up and down repeatedly and slowly, stroking his hard penis.Prompt (pan to, he strokes):
Scene start with a woman posing.
Camera zoom out slowly, then a man's lower body appears from the bottom of the video. His legs are spread, and he is holding his hard penis with his hand. Only his thighs are visible, wearing nothing on the bottom.
Holding his hard penis with pinched fingers, he moves his hand up and down repeatedly and slowly, stroking his hard penis.Prompt (pan to, she strokes):
Scene start with a woman posing.
Camera zoom out slowly, then a man's lower body appears from the bottom of the video. His legs are spread, with a hard penis. Only his thighs are visible, wearing nothing on the bottom.
She leans forward and grabs his hard penis in her closed fist. Holding his hard penis with closed fist, she moves her hand up and down repeatedly and slowly, stroking his hard penis.Prompt (cut to action, he strokes):
No man, not here ✋🏻🤠Prompt (cut to action, she strokes):
No man, not here ✋🏻🤠High Noise LoRA Scale: 1.0 for ongoing and pan, 1.2 to 1.5 for cut to action
Low Noise LoRA Scale: 1.0
Shift: 4
T2V:
Theoretically it works, but I haven't tested it. If you want to test it, keep the same "cut to action" structure as I2V, and start with a low scale on High Noise (0.5).
About HLR + ZCD
This is a fast-learning technique, which makes LoRA less flexible. This can drastically reduce creativity, but it yields stable results while using few resources. Negative effects:
It will probably only work at the same camera angle
High chance of not responding to different prompts
High chance of forcing the original characters of the training video Training only the High Noise reduces this chance, but it can force things that are VERY similar, like the same hair length, same color, etc.
Download
Download High Noise LoRA (on HuggingFace 🤗)
Download Low Noise LoRA (DR34ML4Y_I2V_14B_LOW_V2)
Help me creating more
If you want to help me continue making LoRAs, or if you want me to make a LoRA for you, buy 5000 PlayCoins at Aisha-AI and transfer them to my account (account number 2).
This helps Aisha-AI to stay alive and produce new LoRAs for you all 💜
Description
First version
FAQ
Comments (7)
Got some good results by adding also high noise dreamlay v2 lora.
Wait, you talking you train this lora only on 2 clips? o.O Can you show them somewhere? And... HLR + ZCD?
@forfreelsd368 Adding the High Noise from DR34ML4Y doesn't make sense. High Noise training determines spatial positioning, which means that the same camera position, environment, etc., is almost always used. It's essential for the "cut to". I do intelligent training, almost without a trigger word because I don't want 1000 steps just to fix the positioning and abrupt movements in a single word. That's why my LoRAs are very flexible, the scene can be completely changed most of the time, and the prompt is so precise that if I don't use my LoRA it still works.
But then comes the question: "If the prompt is so good that it can work even without LoRA, then why should I use the LoRA?"
And the answer is the consistency of positioning. Without using LoRA, Wan will be willing to use all its knowledge base to create the initial noise, and that can mean a different angle, a different environment, a different progression. LoRA will only reduce the mathematical chances of creating something far from the training video. One example is my LoRA "discovering a morning wood," where the LoRA's role is simply to maintain consistency in scene transitions, but if you don't use LoRA and request each scene separately, it can still handle it.
@forfreelsd368
HLR = High Learning Rate
ZCD = Zero Caption Dropout
The Learning Rate determines how much LoRA learns from the video at each step. I use a Learning Rate above the recommended level, which causes LoRA to become overfitted (not very flexible, repeating the features of the training video too much). Since I only train in High Noise, this isn't a problem, as High Noise doesn't decide on the face, texture, or other things, but it will make it want to repeat the same overall composition, such as the same positioning, same environment, etc. That's why I can create a functional LoRA in just a few steps, with only 1 or 2 videos (only 1 is necessary).
The secret isn't in the video, it's in the captioning and training strategy, and also in knowing when you need to train Low Noise as well.
The caption dropout is a parameter that decides how many video frames will lose the caption so that LoRA doesn't depend on the prompt to work.
For example, if Wan didn't know what nipples are and I had to create a LoRA teaching that, if I used ZCD (set the caption dropout to 0), I would necessarily need to include the nipples in the prompt. If I do the training with caption dropout at 1 (which is equivalent to 100% of the frames not receiving the caption), I wouldn't need to write the nipples in the prompt, but this could also create nipples randomly without me asking (you must have seen this in image models like Flux and Qwen).
This means that my LoRA needs the prompt to work, which increases compatibility with other LoRAs and also allows it to be merged into checkpoints without trying to force scene transitions and the things that LoRA was made to do.
@AishaAI Thanks for answers. (I don't use translators or dictionaries, so beware upcoming mess, he-he)
About dreamlay high noise - 50/50. Look, I'm the "face consistency" guy. For this it needs to be not just cut transition, it must be smooth transition (when one pic still on screen and another appears smoothly over first pic - wan2.2 knows what to do next) at least, and native wan2.2 comparing with dreamlay high understands that (kinda, so that's why 50/50 now, because I tested it next and it stoped working). Tried different prompts with only your high noise - as you wrote "HLR + ZCD may cause smth smth about face from training or hairstyle", and yes, it does, for prompt with "cut" transition (and I am whole evening testing just one prompt to do it nice, yeah - hobby it is, I think I even wouldn't use much it after I get good result ^_^).
"The secret isn't in the video, it's in the captioning ..."
And that's also. You know, almost nobody shows his training data, so I can't say about "there's nothing special in it" if I can't see it. I can't use cloud services for training or experimenting due to my country, so I tried to train loras on 4070tis and it was possible for training, but not for experimenting. Now I got a better card, so that's why I'm asking - there's really no guides (much of them like bla-bla, here some numbers, do that and will be good).
"and also in knowing when you need to train Low Noise as well."
Everyone: train both.
You: not, if you can use something like dreamlay for low noise.
Me: that's smart, and that's kinda obvious, but when I trained my first lora I even don't thought about that this way (yep, my lora wasn't published because this guy who really trained same stuff same time and I was like WOW, and he did it better, lol).
Anyways, my point is - I wanna got a better face consistency. And yes, dr34mlay high doesn't do much as I knew now. Also tried some transition loras - not very helped (I tried them only on first "cut" prompt, I can write prompt without cut that will helps for face consistency, but I catched myself with interest to do something with "cut" stuff itself without face change, even a bit). If I find something good - maybe even publish it with links to all that I used (but civi likes to nuke everything I post if that not ducks or cats, so...). Your lora is pretty good so that's why I was interested in how you made it (and training two clips because others: 40 clips min, strict manual caption - really boring stuff, I tried).
@forfreelsd368 I said that HLR can cause these problems, but I'll remind you again, I only trained in High Noise, and High Noise does NOT have the ability to memorize a face; that's the function of Low Noise.
If we were talking about a mega close-up of a face on the entire screen, then yes, High Noise would make a difference, but we're talking about a face occupying 20% to 50% of the screen. High Noise isn't affecting facial consistency, and neither is Low Noise in this case, since I'm generating videos with excellent facial consistency. The problem lies in how you're generating the video. I see many people complaining about my LoRAs, saying there's no facial consistency, and at the same time I see people praising the facial consistency and posting videos with my LoRA where the facial consistency is perfect (✋🏻🤠🤚🏻).
I don't even know what to say, people just complain and expect me to guess what they're doing wrong 🤷🏻♀️
I use Replicate's WAN 2.2, they use the standard WAN 2.2 Python script, not ComfyUI... The video is generated in 30 seconds, not because of Turbo LoRA, but because they definitely use several H100s together or a fucking B200 to handle the high demand
Hm, it may not make sense but adding only high dreamlay worked for me, thanks!
Amazing one. mixing with other pov loras it can make a lot of different actions.