Topic: how to avoid jelly flesh genitals?

Posted under AI Art Tutorials

edit: figured out my fps/video length was too much, thanks to the people on discord for the help!

New to AI video generation, using wan2.2

Having an issue where I'll get good penetrative motion, but then it seems like it broadly applies the motion to the vulva making it look like its made of jello/giving it breast like motion. Any tips to avoid this?

Updated

Oarebt

Former Staff

It is hard to help when there is a lot of reason why it could happen.
Model, prompt, configuration, etc!

I will assume that you are using the base wan2.2 model.
Which is not the best for proper NSFW motion.

I would recommend using one of the following merged wan2.2 models.
They automaticly include more NSFW lora/training datas, so NSFW motion are better by default.
Wan2.2 Remix : https://civitai.red/models/2003153/wan22-remix-t2vandi2v?modelVersionId=2770795
DaSiWa-Wan 2.2 : https://civitai.red/models/2272580/dasiwa-wan-22-i2v-14b-or-truevision-or-safetensors?modelVersionId=2769496
Smooth Mix Wan 2.2 : https://civitai.red/models/1995784/smooth-mix-wan-22-14b-i2vt2v?modelVersionId=2513182

But, of course, it will not just make it automaticaly work.
Local AI vid model are still pretty limited.

Feel free to join the Discord, might be easier to ask questions and get answers.

oarebt said:
It is hard to help when there is a lot of reason why it could happen.
Model, prompt, configuration, etc!

I will assume that you are using the base wan2.2 model.
Which is not the best for proper NSFW motion.

I would recommend using one of the following merged wan2.2 models.
They automaticly include more NSFW lora/training datas, so NSFW motion are better by default.
Wan2.2 Remix : https://civitai.red/models/2003153/wan22-remix-t2vandi2v?modelVersionId=2770795
DaSiWa-Wan 2.2 : https://civitai.red/models/2272580/dasiwa-wan-22-i2v-14b-or-truevision-or-safetensors?modelVersionId=2769496
Smooth Mix Wan 2.2 : https://civitai.red/models/1995784/smooth-mix-wan-22-14b-i2vt2v?modelVersionId=2513182

But, of course, it will not just make it automaticaly work.
Local AI vid model are still pretty limited.

Feel free to join the Discord, might be easier to ask questions and get answers.

I'm using DaSiWa's synth seduction v9! did some little tests between it and their other models and it seemed to work best.

Have tried every possible positive and negative prompt I could think of, but nothing seems to be able to fix it. Still pretty new so any prompting tips you have are useful.

Settings wise I've been using refiner + total steps 6 and 10; cfg 1.0 - 1.5 (usually 1.0, tried 1.5 to see if it'd help with my attempts at refining the prompts); sampler is euler, scheduler is linear quadratic; using interpolation. Not sure what other settings could be important to this issue.