So, I know some people won't be happy about this, but after a whole lot of testing, I've decided not to release a low-noise version because it doesn't seem to help and in some cases seems to hurt with this particular concept. For whatever reason, unlike other LORA I've made, ONLY the high-noise is needed. The high-noise gets the characters in the right position where you want them to be without overwriting the styles or whatever LORA. With the low-noises I've trained (numerous attempts now), the WAN tends to imitate the style of the photos used as opposed to the faces/looks you prompt. So, I'm only releasing the high-noise until/unless I can train a good low noise, but like I said, it really doesn't need a low noise (see samples).
Description
This is not actually the first version but it's the first version I'm posting
FAQ
Comments (21)
High noise only can be legit for motion only LoRAs. Low noise is for detailing. If you only want movement, not adding extra details (especially for things the base model already understands), it actually makes more sense to run just the high noise. A low noise would just be pushing the generation towards the specific details you used in the training data and thus less flexible.
Ahh. That makes perfect sense why it doesn't work well with this concept. Because it overwrites details when all I really want is the position + motion. If enough people demand it, I MIGHT take another shot at the low noise. Right now, my only evidence is "trust me bro," but it really does seem to be 100% unnecessary for this type of LORA.
So to make things short did i understood well that for pose concepts like these that wan dont understand at default ( two subjects in intimate position doing their thing ) only the High noise lora is needed ? Like i am planning to redo my reverse headscissor lora for 2.2 with better dataset, and hearing this i wonder if i would need only to train high noise if i just want this pose concept to be recreated on wan environement. Just woman on top of man squeezing him with thighs and feet crossed. I need to know this
@magmirant0011 Wan is not great at nsfw bodies or nsfw motion, so it can make sense to have low noise for motion LoRAs that also have genitals in the shot. If no genitals are seen and it is purely motion, you probably don't need a low noise LoRA. That said, like most things in life the answer is usually "It depends."
Yeah but this is why it looks like a man is just sticking his face in her butt and her cheeks don't spread of move.
@magmirant0011 It's funny you should mention that because I actually USE your LORA a lot and I only use the high noise lol.
@CAldonni I think that is just based on the seed. I just generated another one and they clearly move. I'll upload it right now.
@magmirant0011 just always train the High lora first, and then when testing you can see if you need the low lora to properly add details
Is there a trigger phase?
Not in this one, no.
Reversed the genders and used a female character lora and it works great.
I tried that and the motion works, but it makes the cock that is hanging down dissapear. Any solution?
Often for the loras that are focused on movement on character and skin types that are already present, the high is all thats needed. If for example your training really focused on the tongue and the genitals, a low would likely be needed to properly resolve those details. But in this case, letting the vanilla work is all thats needed
That's exactly what it seems to be. I just worried because, people being people, I know there will be some who are very mad/upset that there's no low noise. But I just don't see a benefit here.
Thank you for this awsome Lora! Only High seems to work perfect! I'm surprised.
Can i ask if you use only images to train this Lora? And how many images? Or a mixture of short clips and images? Differnt image sizes? Sorry for all the questions.. :) I want to make a femdom Lora but no idea how.
For this one in particular, it actually had WORSE performance with videos. This was the opposite case from Wan 2.1, where I used videos to get the best results. I believe it's because WAN 2.2 is much better and understanding motion via prompt and therefore all it really needs is a kick to get the characters into this position.
@CapAndABull Thank you for the insight! It actually makes sense. Video are essentially just made of images. Can i ask how many images you used for best result? When asking Chatgpt it says 120-150 images for training concepts, and include "screenshots" of a few clip where you grab only 5-6 frames (Imagine low fps) - to help train movements.
¿Por qué querría alguien tocarle el trasero a una muchacha? Debe tener tantos bichos...
Se llama no ser gay
Works great on I2V too, thanks!
what is the recommended strength and other settings? I'm getting some unnatural lack of motion / stiffness in the animation.
Details
Files
Wan2_2_Ass_Worship_HighNoise.safetensors
Mirrors
Wan2_2_Ass_Worship_HighNoise.safetensors
Wan2_2_Ass_Worship_HighNoise.safetensors
FemdomAssWorship_HIGH.safetensors
Wan2_2_Ass_Worship_HighNoise.safetensors
Wan2_2_Ass_Worship_HighNoise.safetensors
eat_ass_HIGH-NOISE-ONLY.safetensors
Wan2_2_Ass_Worship_HighNoise.safetensors
Wan2_2_Ass_Worship_HighNoise.safetensors
FEMDOM Ass Worship T2V+I2V HIGH - Wan2.2.safetensors
FEMDOM Ass Worship T2V+I2V HIGH - Wan2.2.safetensors
FemdomAssWorship_HIGH.safetensors