Stroke it honey 😏
Yes sir 😇
Training
200 steps on high noise
2 edited videos (40 frames each video)
HLR + ZCD (These acronyms were created solely to confuse you)
Lots of confidence
$3 (✋🏻🤠🤚🏻)
Problems with facial consistency
I don't train any of my LoRAs using Turbo LoRA, so you should avoid it too.
It's still possible to get a good result using Turbo LoRA, but for that you'll need to increase the number of steps or slightly increase the CFG Scale.
If you don't know what the CFG Scale does, it's basically the parameter that defines how much the prompt will be followed.
CFG 1 tells the WAN that it can follow its instincts, while the WAN's default without Turbo LoRA is CFG 3.5, which gives more importance to the prompt, and that's why you lose facial consistency.
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):
She holds his hard black penis with a closed fist, moving her hand up and down rapidly, stroking his hard black penis.Prompt (cut to action, standing):
No man, not here ✋🏻🤠Prompt (cut to action, kneeling):
No man, not here ✋🏻🤠This LoRA become very flexible, you can completely change the scenario and it will still work
High Noise LoRA Scale: 1.0
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.
ZCD = Zero Caption Dropout, this means that the training was conditioned by the caption. This causes the LoRA to be activated only when this exact (or very similar) prompt is used, increasing compatibility with other LoRAs and allowing it to be incorporated into the checkpoint without trying to force the execution of what was learned.
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 (14)
I've been able to get your loras to work, but this one is fighting me. The character changes vastly from the latent
There's not much I can do. You didn't say how you're generating the image, and I obviously do consistency tests before posting all the LoRAs (especially since I sell the generated videos, so it needs to work properly).
All the videos here were generated on the first try, with random seeds.
If you're using Turbo LoRA, that's the problem, and if you're not, it's something else, because it doesn't make sense for it to work perfectly on the standard WAN 2.2 and not with you. So if you want help, tell me how you're generating the video and maybe I can tell you what's causing the problem.
@AishaAI I'm genning on Comfy. I copied settings I've used for your other loras as a baseline and something isn't working for me
- I'm using base I2V FP8 for high and low for the checkpoints
- I've tried different combos of lightx2v for high and low. Killing lighx2v for high seems to help out
- I've also tried wan.2.2 lightxv2 to test it out
- Your lora in high @ weight of 1, dreamly v2 in low @ weight of 1 (and less to try it out)
- I'm using Wan MoeKSampler and I've tried 3/5 and 4/4 in high to low shift
I think I'm missing some obvious here? I've run and test many loras and can usually get them to work. Just something isnt playing nice on this one and I can't figure it out
@MankyPoodle You said "Killing lightx2v for high seems to help out," so from what I understand you stopped using Turbo LoRA only in High Noise. That doesn't make sense to me, but I don't use ComfyUI so I don't even know if that's possible.
The big issue is that if you need to stop using Turbo LoRA, it's in Low Noise and not High Noise.
High Noise is only responsible for the beginning of each frame, deciding the positioning and abrupt movement; it doesn't decide faces or other details (that's why I only make LoRAs in High Noise). Low Noise is responsible for maintaining facial consistency, making micro-movements, and texture. My recommendation is that you simply stop using Turbo LoRA, because it's OBVIOUS that it will give up something to create the frames faster. If you don't want to stop using Turbo LoRA, gradually increase the CFG Scale, as it controls the importance of the prompt in relation to the "instinct" of the Wan 2.2 base.
@AishaAI What do you consider a turbo lora...Lightx?
@MankyPoodle Yes, that's Turbo LoRA.
@AishaAI gotcha...i was think you meant a turbo/merged check point.
I'll do some testing without then. It's still odd I got your other loras to work with my "standard" setup with turbo.
I'll update in a bit
Still no luck in removing lightxv2 in high and low both. I cranked steps up to 12 and 16.
I'll just note this one as something not playing nice in my setup and I'll continue to enjoy your other Loras :)
@MankyPoodle Send me the prompt and maybe even the generated video
@AishaAI how can I send?
@MankyPoodle the prompt you can send here, and the video you could host on a temp file host like https://tmpfiles.org/
@MankyPoodle Do this trick:
frames: 33
aspect: 9:16 (input image I use layer utility node with crop and lanczos)
width: 351
heigh: 624
High pass
Power Lora Loader:
- Wan21_i2v_14B_lightingx2v_cfg_step_distill_lora_rank64 at 0.20
- bbc_handjob_wideview_000000200_high_noise at 0.80
- KSampler (Advanced) steps: 4, cfg 0.8, lcm/simple, start 0 end 2
Low pass
Power Lora Loader:
- Wan21_i2v_14B_lightingx2v_cfg_step_distill_lora_rank64 at 1.00 (You can up if you want but start with 1.00)
- Wan21_CauVid_14B_T2V_lora_rank32 at 0.40
- DR34ML4Y_12V_14B_LOW_v2 at 0.30
- KSampler (Advanced) steps: 4, cfg 0.8, euler/simple, start 2 end 4
I use frame_rate at 16 so it looks way more real.
Have fun!