For on-site generation, I recommend only using models designed for WAN2.1 or earlier. It does not appear that Civitai can get 2.2 models to work. You can use the Universal model with both high and low noise steps, but if you want the best quality, I included both versions as well.
This LoRA is trained on conversational language. You can describe your scene in plain English or Chinese, and it should adhere to your prompt well. Below are some inspirations for you if you need it.
Positions:
On her back
Kneeling
Perspectives:
POV
Enhancement phrases:
She bounces her body up and down
She bounces her breasts up and down with her hands
All showcase videos have the workflow metadata embedded. You can download the video and drag it into ComfyUI to replicate them with their prompts, seeds, and settings.
Description
Includes better nipples, under clothing variation, and better adherence for side-view angles. Additionally, you won't have to be as overly descriptive with your prompts to keep the penis between her breasts. The LoRA now understands that Titfuck means "penis between breasts" usually.
This version was trained using the following configuration and dataset settings:
26 videos 4-5s long each. 256x256. Downscaled using bicubic to minimize pixelation
200 Epochs, each video is processed once per epoch.
Learning rate of 2e-5
Captions had reduced redundancy (removal of "penis between breasts" and other overly descriptive elements) to allow the LoRA to increase association with "titfuck" instead of other tokens.
Introduced more side-view videos, videos that included better nipple detail, and titfucks with clothing (bikinis, sports bras, etc)
FAQ
Comments (37)
Awesome Loras, great work
Stacking loras definitely diminished the movement on it, but still captures it well enough. Great work.
Some LoRAs that affect body movement are going to be really hard to combine. They both likely affect the same vectors. Other LoRAs (Stylistic, or enhancements) like Move Enhancer or Cyberpunk should work fairly well.
@ComfyTinker The full release version does great at preserving the movement while also not impacting faces too. Great work! Mind sharing your training settings? LR, steps, batch #, frame count and resolution, that sort of thing? Might help me with some complex movement concepts I'm working through with training.
@KinkMaster 256x256 videos at 24fps. They were downscaled, but I made sure to use bicubic downscaling to avoid pixelation. They were all manually captioned fairly in detail. I captioned individual facial details when their faces were in frame, and I included a lot of videos that kept the face out of frame, as well.
Total epochs - 200
Total input videos - 65
Total steps - 13000
Loss rate - 2e-5
Gradient accumulation steps - 1
No repeats per epoch
@ComfyTinker Thanks a ton! Very helpful!
lower the guidance for more movement in any scenario
increase the strength to 1.2 to get the motion back when using a character lora
Where is the diffusion model from your workflow from?
https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/
I use the official comfyorg repackage, but these LoRAs should work with others.
@ComfyTinker Do you have a link to the comfyorg repackage? I'm using a 1 click install of comfy called windows portable.
@crazymanjj I linked to comfy's official example page. you can download it from.
@crazymanjj do u get the workflow?
@forbiddenkingmaker1230680 It's in the metadata of the videos.
@ComfyTinker where to find the metadata
A noticeable improvement from 0.9 to 1.0. No more movement loras necessary and much less praying to RNGesus for a coherent result.
I plan to revisit my spanking LoRA with what I've learned.
It works really well, great work. It doesnt seem to affect char lora too much, which is nice. Did you remove faces from the training data?
Also, in your Tips you have a typo "She is titfucking his penis between here breasts."
Thanks for the correction, I've updated the typo!
As for faces in the dataset, I used a bunch of videos with the face cropped out, or when the face was included I made sure to describe them with.
Ex:
"She is brunette woman with hair is loose waves with thick eyebrows, blue eyes, thick lips, and a seductive expression looking up. Her hair is in a messy bun"
When I drag the video to ComfyUI online, it says no workflow file is found. I tried 2-3 videos and all say the same. What can I do?
Not sure how ComfyUI online is set up, but I would just set up a standard HunYuan workflow and copy the prompt I shared in the description, then.
@ComfyTinker Okay thanks, I am new to all this stuff. I shall try your method.
Where you training WAN loras? Is it really that similar to Hunyuan?
I use diffusion-pipe. Yes it is very similar to HunYuan
@ComfyTinker For the upcoming 14B WAN variant, is that compatible with only the 14B t2v model or it would also work with the 14B i2v model?
you just drop the wan models into unet and vae folders and youre good to go?
@K3NK use the wrapper https://github.com/kijai/ComfyUI-WanVideoWrapper . From my testing Hunyuan is way better than the 1.3B model. The 14B model doesn't work for me atm
I haven't seen a unet but I was able to run both the 1.3b and 14b WAN models after updating ComfyUI and using the workflows from this guide https://comfyanonymous.github.io/ComfyUI_examples/wan/
@OnlyGANs 14B LoRA models are cross compatible for both I2V and T2V from my testing.
@WhatTheGuy Hunyuan has 13B parameters, so it is expected that the Wan 1.3B model will be worse. But I suppose that Wan 14B should surpass Hunyuan, at least in motion quality. And if there are no problems with censorship, Wan will blow up the community again, as it did with Hunyuan.
@Aivanjo yes the 1.3B is the smaller model, but basic concepts like a blowjob work with base hunyuan and isn't possible at all with base WAN. So with hunyuan it's just finetuning what the base model already knows. WAN has never seen it. So it's a entirely new concept for WAN.
I'm not sure if the bit better model will produces better sex scenes when it entirely depends of the finetuned data. While with Hunyuan the power of the basemodel already knows the concept and you are just perfecting it.
@WhatTheGuy Yes, that's what I meant when I talked about censorship. I haven't tested the WAN model myself yet. But if it doesn't allow generating sex and nudity without additional lore, in other words, if the model's dataset didn't contain such videos, then attempts to further train the model won't be very successful.
@Aivanjo you can generate naked breasts at least =) But always with underwear. Hm maybe this will become the new Flux of video models ;P
@WhatTheGuy Another model to test the community's decensoring skills. Sounds exciting 😁
Haven't tested myself, but I'm hearing SD2.1 here...
Can a video be made by just writing one Titfuck
You could in theory, but HunYuan typically needs more description than just this. Also you can titfuck a dildo, change clothes, change hair, and a bunch of other variations, so just the word "titfuck" might allow too much creativity and wouldn't get you where you need it.
"A naked woman titfucking a naked man." would be the bare minimum, and would still require some prompt-fishing.
"A naked woman titfucking a naked man. She is kneeling between his legs." would be my recommended starting point.