As requested in the I2V version, here is the same dataset/captions made for T2V wan. It works great out of the box for non-nude, but it seems to struggle with getting accurate female genitals. I'd highly recommend using HearmemanAI's vagina lora with it to get more detail in those lady bits:
https://civarchive.com/models/2109996/wan-22-pussy-and-anus-lora?modelVersionId=2387016
Prompt is exactly the same as the I2V:
A woman places her buttocks over the camera, obscuring the view. The final frame of the video is a close up of the woman's buttocks and crotch.And if she's naked add:
The woman is bottomless and her vulva and anus are visible. If you're using the pussy lora change vulva to vagina. I find it doesn't move as quick as the I2V one so adding:
The final frame should be nearly entirely black.Can make her scoot her boot a little faster.
Description
FAQ
Comments (6)
Hmm. Love the concept, but it seems to be exerting WAY too much control over the face or something. Skin texture is super plasticky, and it absolutely OBLITERATES any character loras. Tried lowering low noise strength (since this seems like mostly a high noise/movement focused lora especially), but no luck. LOVE the concept, just not really working like any other T2V lora I've tried. Like I said, may just be me.
Thanks for the feedback, I'm honestly a totally new to T2V WAN loras so i just used the same dataset as the I2V version. I'm wondering if it needs to be re-captioned or better captioned for T2V. I've been meaning to re-caption the whole dataset anyway so maybe i'll do that for both over the next week or so.
@imb101 It typically means the LoRA was over trained using to many epochs or to many repeats. Don't rely on the loss rate on the tensorboard. The loss rate is not a good measure for overfitting. You need validation testing for that and most trainers do not have validation testing for Wan.
For people using the LoRA, to fix this issue it typically means using a lower strength on the low noise model. The low noise model has a greater influence on the 'visual' of the output, high model is motion. Not sure why the poster is still having an issue - may need to lower more or look at how the split on the high/low samplers (sigma shift, boundary, etc).
@Kierkegaard420 Ahh interesting, thank you. So i'd cranked up the low noise model to 26 epochs, with 62 videos in the data set, because i was struggling to get the right details on the skin but maybe i went too far. I've got every 2 epochs saved for both high and low so can try testing it out later today. It's interesting that I2V is more forgiving for over training, but i guess it does have a baseline to start with the image.
Hi! o7
I2V model was nuked?
yes, the mods told we aren't allowed to post any models that move someone from a non-sexual context to a sexual one. I'm fairly sure that's about 80% of the I2V loras but i've stopped trying to make sense of these things and just assume the 'I just work here and i dont make the rules' applies to most mods :)
Anyway, you can still get it from my huggingface page, which is linked in my civ profile. Other loras i can't post here i will post there. Civ archive also has a back up.