you can't spell LORA without ORAL ;)
Note
This LoRA was trained exclusively using images so the motion will be completely dependent on your prompting. Terms like "her head bounces up and down" and "Dynamic Movements" may help to get the desired effect.
My first upload to CivitAI! Trying to do my part to accelerate this new era of content generation.
I noticed a surprising lack of NSFW loras so I thought I'd take a crack at it! Honestly Hunyuan already does a pretty decent job with scenarios like this, this will just add some additional guidance and hopefully give you some more consistent results.
This is actually the second rev of this lora which will try to guide your generation to have a woman performing oral sex on a man. I've had some success including that phrase "woman/man performing oral sex on a man" but it's still a bit of a roll of the dice.
I haven't tested gay content much as it wasn't part of the lora data set but it seems to be possible.
My settings:
(Optional) Clip: ViT-L-GmP-SAE-TE-only --- You can disconnect the Load CLIP node if you don't want want to use this
Lora strength: 1 - 1.2 (you'll have to fiddle with this, if it's too high it may cause the video to become blurry)
Model: hunyuan_video_720_cfgdistill_fp8_e4m3fn
Dimensions: 512w 320h
num frames: 101
steps: 15 (you might wanna pump this to 30, but I was going for speed)
embedded_guidance_scale: 12
flow_shift: 7
denoise_strength: 1
enable_vae_tiling: true
temporal_tiling: 64
spatial_tile_sample: 256
auto_tile_size: false (honestly you can just turn this on since I'm using the default settings anyways)
I trained this with about 40 images downloaded from various free galleries mostly of Asian women. The training was ~ 1000 steps and this is the 10th epoch. I don't know good settings for training a concept yet, so if anyone has some advice I'd love to hear it!
Description
First upload
FAQ
Comments (12)
Just an additional note. Even though this is a POV LoRA it actually does seem to help with other angles like from the side.
Also, it seems like if you combine this with other LoRA (character lora for example) you may have to lower the strength as it introduces bluriness. There might be a way to offset this with lowering the guidance scale but this needs more testing.
good result i always train 20000 steps may be it is unnecessary...
Holy moly that's a lot of steps! Is that specifically for concept? I was reading that for character likeness ~ 1000 / 10 epochs is the sweet spot but I haven't heard much around concepts. Thanks for the tip in any case!
BINGO! 69 downloads! I hope you are all having good success with the LoRA! Feel free to share your creations or any questions you might have. I hope to try and train a new version soon using video in the dataset to get better motion.
thank you sir
Hi. What type of hardware is needed to train a Lora for Hunyuan?
Lots of info here: GitHub - tdrussell/diffusion-pipe: A pipeline parallel training script for diffusion models.
I still don't understand the training process well, and it's also not officially supported, but you'd probably want to have at least 16gb VRAM when training using images (which this Lora is trained on). With video (which is something I am just starting to look at) you're going to need 24GB of VRAM. During a test training I was using ~ 22-23 GB throughout the process.
the other weird thing is the tool doesn't work natively in Windows, as mentioned in the linked page, so you'll have to have a Linux rig or be willing to set up WSL in your Windows machine. It's definitely NOT a user friendly process just yet, but once you've got it set up it's pretty easy to run new trainings.
@Fresnelz How long did it take for this lora training in local?
@cnk0101 On my 4090 about an hour or so, maybe a little more