My first release of this, trained on a nsfw dataset (primary missionary girl focused) but is somewhat flexible. your mileage may vary when using tags beyond pretty ordinary nsfw.
I2V preferred, prefers 2.5D/3D. Cog1.5 2D is usually not a good result, even with 2d in the dataset.
best resolutions for inference (can do landscape or portrait):
720x480
960x720
1216x896
Additionally, trained on 88 frames, so you can go a little longer as well
can do others, but these are the best for results imo, you might get no motion or random motion otherwise.
preferred strength:1.0, but can go between 0.9 and 1.1+
Please use responsibly
nsfw is the trigger, but be sure to add some more descriptors like example:
a girl is having _______ sex with man out of frame, his penis going in and out of her pussy repeated motion(slow motion) might also help, but facial expressions are recommended also (moaning, etc.) the more descriptive you are the better the results. cog uses t5 so it likes more tokens :D
Description
First training run, I plan to expand the dataset to include some other scenes, and try to caption towards speed of motion as well, but thats gonna take time and I could really use some support :3
FAQ
Comments (21)
So good bro!
Any entire prompt examples suggestions?
other than whats in the description, be sure to detail the scene with like the background or other important details to the image composition. Your input image is more important than the text. it does prefer pov missionary/cowgirl as the data is biased towards this.
This is quite amazing! What kind of hardware did you use and how long did it take to train?
4090 over the course of multiple days.
@CubeyAI Is there a guide that you would recommend if it is not too complicated? I would love to train one myself with my 3090.
@divanova there isn't a guide, but you can start here https://github.com/a-r-r-o-w/cogvideox-factory this will get you running but its far from painless to get running
How do I use it? Is with some specific extension?
the only way I'm aware of is through comfyUI using the cogvideoXwrapper custom nodes.
Any plans for original CogVideoX-5B? not 1.5.
not atm, the newer models are getting loras so I've been toying with that.
another request, any comfyui workflow examples?
the workflows are the same as provided by the comfyUI wrapper made by Kijai, I did not make any significant changes
@CubeyAI Can you upload a demonstration project?
i use this lora in cli_demo.py , error is "ValueError: text_encoder is not found in self._lora_loadable_modules=['transformer']"
I've only used this through ComfyUI, I would not be able to assist you on this.
In which directory should I put this file?
i installed cogvideoXwrapper custom nodes.
why the face always crash ？
Probably because it's not the focus of this model. It's focused on the motions in the female body while being thrusted over and over. Just guessing.
Can you please make separate loras for mmd/3d/live2d/2d hentai?
Because they are really different and so they didn't mess up eachother .
If you need high quality videos recommendations i can try to give you a list of high quality artists
yo comfyui user here and im very new... any tutorial on how to use this model...?
Good lora,can author give exmple photos with workflows?Thanks!