Okay babe, you can nurse 😏
Thanks mommy 🤤
Attention ⚠
I wasn't able to complete all the planned steps, so the LoRA isn't VERY good, but it's working. If the generated video isn't following the prompt, increase the CFG Scale of your workflow a bit (I generate the videos via the Replicate API, so I can't increase the CFG Scale of mine 😞)
Training
260 steps on high noise (the goal was 320 steps, but we ran out of money 😞)
1 edited video (80 frames)
HLR + ZCD (These acronyms were created solely to confuse you)
Lots of confidence
$4 (✋🏻🤠🤚🏻)
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
As you know, Civitai likes to do some pointless censorship, and that's why I post everything in my HuggingFace repository.
Visit this LoRA on HuggingFace and see all the prompts and examples.
I2V:
Prompt (ongoing):
While holding the man's hard penis with her closed fist, she moves her left hand up and down rapidly stroking the hard penis while the man licks her nipple.Prompt (cut to action):
No man, not here ✋🏻🤠High Noise LoRA Scale: 1.3
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
Training is over 😎
Check the most recent version
FAQ
Comments (29)
Please upload high and low separately for archiving efforts. ZIP files deprive the loras of site metadata.
bro the training is running NOW, this is why not have LoRA yet
@AishaAI Oh my apologies. I didn't mean to offend. Just good practices.
the download is just a video no lora ????
Maybe you should read the post, maybe... just a hint
Bro, I've noticed that most people who train LoRa to solve the face consistency problem do it by tracking faces and then blurring them. Modern video editing software makes this easy, like DaVinci Resolve or After Effects, which is much more effective than fixed-position blurring.
I only blurred it to post it here, the video I used in the training isn't blurred
LoRA only lacks facial consistency when the person training didn't know what they were doing. I only train with the High Noise Model, which makes it impossible to memorize the faces in the training videos. Wan knows how to maintain faces in scene transitions, even in completely reconstructed scenes, so there's no need for LoRA for that, just a good prompt and decent Low Noise Model training so that no one's face is memorized and forced.
@AishaAI Okay, looking forward to the results of your new Lora.
@jevencheung I'm having trouble creating an image of a nursing handjob already in progress to test the "ongoing" prompt, since I can't post the "cut to" version here. But the LoRA is ready and will be available for download soon
@AishaAI This is a great scene theme! You always have some very creative ideas.
As you guys know, I don't use Turbo LoRA, so it might not work properly if you do. Also, I haven't been able to finish all the training steps and will probably update (redo) this LoRA later. Please read the post before asking questions, that's all, thanks 🤠
this model would be awesome if it had facial consistency
Bro, see live tests
ur doing a great job but please mate switch to training for turbo majority of people uses that
I make those LoRAs for me (to run at aisha-ai.com), and I don't use Turbo LoRA. You can use Turbo LoRA and still get good results, just increase the CFG Scale a little. If you don't know what CFG Scale does, just ask any AI and you'll understand why using Turbo LoRA makes you use CFG Scale at level 1.
我尝试了一下这个lora，令我吃惊的是它的人物一致性比绝大多数都要好,做的非常棒！
When you When you know what you're doing, things work out well. Thanks for sharing 💜
A picture of a man and a woman at the beginning, how to write the prompt words for the scene switch?
Simply replace "man" or "the man" with "he" or "him"
How to write a prompt for cut
"Uncensored prompts
As you know, Civitai likes to do some pointless censorship, and that's why I post everything in my HuggingFace repository.
Visit this LoRA on HuggingFace and see all the prompts and examples."
@AzAsAsyn Thank you, friend, you are a true friend, my friend
What... wan loras can actually be made with only 1 vid in dataset?
Only if I don't want to teach several different angles. I like to take full advantage of the enormous learning that Wan 2.2 already has, or pieces of LoRAs that other people have already made.
In this example, I use the Low Noise from DR34ML4Y, which required many more videos and many more steps to get good. If I were to train my own Low Noise, for example, I would need to include many more videos and train with MANY more steps to be able to learn, so I try to take advantage of what Wan already knows, what the community has already trained, and the scene control of High Noise, which usually gets good with very few steps. If you don't know the difference between the two, High Noise is the model that decides the overall composition, where the people are, camera movement, arms, legs, etc. Low Noise is the model that does the refined texture of objects and also the precise movements of the video. So if High Noise does an easier job, it learns faster, and that's why these LoRAs of mine are possible and work 🤠
@AishaAI Wow, I wasn’t aware of that. I had a few LoRAs I wanted to make, but I didn’t have many clips to use for training. Now that I know I only need one clip, that changes everything. So thanks.
I have another question: what is HLR + ZCD?
@naomixlove
HLR = High Learning Rate (Learn very fast, around 3 epochs)
ZCD = Zero Caption Dropout
Caption dropout = some steps will be learned without parts of the caption, this will activate the LoRA even if you don't write the correct prompt. For a LoRA of style, it is good, for a LoRA of concept it can be very bad
@AishaAI ZCD sounds interesting. You know if it is possible to apply that in AI Toolkit? Or would I have to do that manually by repeatedly pausing the steps and putting in another caption?
@naomixlove yeah, I use that. Just set caption dropout to zero