v1.4: Recommended if you want to operate with simple prompts. However, there is a problem with controlling the face and expressions. v1.5: This model improves on the face and expression issues. However, the prompts are difficult, and it is likely to be less overall good than v1.4. This model is for WAN2.2 5b. It operates with VRAM up to 6GB and can create a 720x480, 129-frame video in under a minute on a 3060.
The Lora model has been updated to v1.4 and has been improved to perform better on all keywords.
This Lora only supports the anime style.
Includes the main tags nsfw_5B, as well as sex, missionary_position, doggystyle, cowgirl_position, handjob, blowjob, footjob, paizuri, and cumshot.
For prompts, please refer to the showcase.
It's better to write about the character's movements, the object's movements, or after how many times...
For ejaculation scenes, I think you'll get better results if you write down the start, intermediate drawings, and the end.
If you have a better prompt please share it!
Description
I'm planning to increase the dataset a bit more.
FAQ
Comments (7)
2 questions.
Is it really censored jav style or its just for the samples?
And theres only one file, do we use it as high noise or low noise?
Thanks
This is a Wan 2.2 5B lora, so there is no high or low noise. It will not work with Wan 2.2 14B.
None of the previous versions have been censored, so would be weird addition for 1.5 so I feel confident before checking it out that it's not the case.
Still has that "moving paperdoll" look rather than actual animation.
This isn't a fault of the lora however, it's just how all 2d loras turn out. Idk why all video generators look at 2d and immediately go either "Scribblenauts animation!" or turn it into 3d.
At this point, I think the only real solution is for someone (not OP, obviously) to train a full base model on animation only. Maybe someday they could be merged with a live-action base model to make a fully featured one, but I think we need that 2d base model first. I've yet to see a lora meant for 2d/traditional animation actually look like animation.
Thank you. You're right, incorporating MMD models and live-action videos might be a good idea. When upgrading from v1.4 to v1.5, we experimentally trained with a small dataset, using only animation. We plan to increase the dataset when upgrading to v1.6.
@diffusion078746 Nah, I'm saying to exclude 3d and live action altogether. There's so much of it in the base model already it makes itself present when you're trying to do 2d videos. You either end up with the paperdoll style that looks like my link, or the 2d animation turns 3d. In either case it never actually looks like traditional animation.
I don't think a lora will be capable of fixing this, but perhaps an increased dataset with hand-picked examples will work.
@anon4324 It's an anime style completely without CG or MMD. I see, it seems difficult, but I might try it someday.
Here the list of prompts.
nsfw_5B, twerking, A girl with a hand on her knee bend over and shake her pelvis up and down.
nsfw_5B, kissing, A girl and A viewer kissing each other.
nsfw_5B, grinding, A girl on top of viewer rub her pelvis back and forth.
nsfw_5B, humping, A viewer standing behind her and moves pelvis back and forth.
nsfw_5B, sex, doggy, A girl and a viewer have sex. The viewer moves pelvis back and forth and insides his penis into her vagina.
nsfw_5B, sex, missionary, A girl and a viewer have sex. The viewer moves pelvis back and forth and inserts his penis into her vagina.
nsfw_5b, sex, on side, A girl and a viewer have sex. The girl moves pelvis back and forth and inserts his penis into her vagina.
nsfw_5b, sex, reverse cowgirl, A girl and a viewer have sex. The girl moves pelvis up and down and inserts his penis into her vagina.