The Hunyuan version of my SDXL lora which can be found here.
Again, this fetish depicts a motion which can be quite strange so consider it very experimental. Trained it locally so the dataset is of course limited. Still, I'm quite happy with the results and considering it's "one of its kind" you should be too ;)
In any case, I will probably iterate over this so follow if you are interested in getting notified when a new version drops.
Description
FAQ
Comments (11)
Thanks! I'm always on the lookout for more Hunyuan video loras.
The quality of this is really impressive. The subject, um, less so. :)
Would you perhaps consider applying this level of quality to an 'undressing' lora? You can have all my 166 buzz.
Us Hunyuan enthusiasts are lacking in the undressing department.
Of course, only once you've finished catering to the penis sniffing crowds. :)
Thank you! appreciate the comment. And yeah, penises on noses is kind of niche... it either floats your boat or not :)
Edit: sorry didn't read the full comment, haha, penis sniffing crowd good one. Undressing lora uh? Yeah let me think on what the dataset would look like. The issue with it on top of my head is that it fully relies on motion, and that requires video training data, which is obviuously far more gpu hungry thatn this one. In any case I will give it a thought and see if I can experiment with that.
Can you tell me how to train locale? i tries using oneshotlora.com but somehow that wont work well
Hey! I use OneTrainer. You can find it on github, documentation is there and you have plenty of tutorials online. Cheers!
@sparkfun i installed it but i am a bit of noob no idea how to do it , any tips? or a noob explanation?
@erik28173 sure, you first need to get familiar to what you are aiming for and prepare a dataset for it. In the case of this cock sniffing Lora version the dataset is exclusively images of penises pushing noses on different perspectives and characters. The dataset does not contain videos because motion is not necessarily part of the videos you want to create with it - and of course is much (much) more expensive computationally, so I sacrificed "variety in motion" for "effect consistency of the penis pushing a nose". Then you need to prepare a description for each image and try to capture what the image looks like to a "text-to-vision" algorithm like Hunyuan, so focus on describing how the motion affects the physics of whatever there is on the image and completely ignoring the character themselves (unless you are training for character likeness of course).
I hope that gives you an idea of the dataset preparation which is by far the trickiest part. Now for operating the actual OneTrainer software it's pretty straight forward, you can select Hunyuan video default setting, create a concept with your dataset and train it. Trial and error is your friend. I encourage you to search around for Onetrainer tutorials for specifics as this is a very broad topic.
Cheers and enjoy.
@sparkfun Thank you very much for your good explanation i will see if it will work for me ,and hope you also make more models for hunyan
The uploaded LORA is not in comfy ui lora format. Please convert it from musubi format to comfy.
There are nodes for Comfyui that load musubi loras no problem. Actually the OP gallery is generated with Comfy.
Edit: That being said I’m not super familiar with the musubi format specification. I trained it with OneTrainer.
This is great! I love this. Hoping there will be a similar lora for the lips <3