Further finetune of POV Oral | Hunyuan Video LORA,
trained on 748 samples in the dimensions of 512x512x33, 640x384x33, 384x640x33, collected from around 55 video-clips collected from realbooru.com (cock_worship, deepthroat, cock_hungry tags), all trained with the text-prompt "A woman performing oral sex."
Trained on 4 gradient accumulation steps for 9 epochs.
512x512 videos seem to perform the best, vid2vid works nicely.
Description
FAQ
Comments (27)
good shit
Now we just need a HJ Lora
I've only just started playing with txt2vid using this: https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/ Can I use this LoRA and if so, how?
I see comfyui now can do hunyuan out of the box.
put the .safetensors lora file into models/loras/ and then put LoraLoaderModelOnly after Load Diffusion Model. Under Model Name select oral_e9.safetensors and you should be good to go after connecting the lora loader node to the rest of the workflow.
low rank adapters are basically little modules that you add onto existing models. you can select how strongly you want it to be added and you can add multiple different loras too.
Adding the lora node in comfyui: Right click in your workflow. click "add node" , click "HunyuanVideoWrapper", click "Hunyuan Lora Select". connect this new node to "Hunyuan Video Model Loader", it has a lora Input.
@dgaa Thank you, as far as I can tell it is now working perfectly.
Sorry, my text refers to Kijai nodes
I am new to Comfyui and was not able to connect this Lora. Is there a workflow which I can download?
CUrrently I am using this workflow:
https://blog.comfy.org/p/hunyuanvideo-native-support-in-comfyui
@CyclopsGER you can do it i believe in you
This works really well. I hope you'll do more lora. Can't wait to test more sex positions.
Like if you need some ideas: suspended congress. The base model will always return the girl facing the viewer, but I struggle with it facing each other, chest to chest. Sometimes rarely I get it to work, but it often has clipping.
Maybe this helps to easily find trainings data:
This works great, would you mind sharing what you had set the learning rate to on your training?
just the standard values from diffusion pipe
type = 'adamw_optimi'
lr = 2e-5
betas = [0.9, 0.99]
weight_decay = 0.01
eps = 1e-8
it feels a little low though, 3e-5 could possibly be better
Actual good and versatile lora, you should do the same for vaginal sex.
Is it a way to use that with A1111 ?
a1111 hasn't been updated in months bro just take the comfyui pill
Seconded, gotta move on to FORGE and Comfy. Forge is A1111 but better and (faster for SD and XL) and Comfy for all the bells and whistles now is a no-brainer, it's not as daunting as it looks.
@admiral_underpants yes i have Forge installed through Pinokio, just dont use that for now because of a bug with LORA previews img, i will try again to switch im bored about of A1111 stagnation and limitations
This is why I recently moved to comfy. Def worth learning. The UI might look intimidating but it's not that hard to learn especially when you know how A1111 works. I instant noticed quality improvements in my work. I just use Forge and ComfyUI now. I'm sure there'll be support for it eventually... and I ain't waiting for all that
bf16 VAE gives: 'Error(s) in loading state_dict for AutoencoderKLCausal3D:' and a litany of encoder errors, what's the right VAE for the fp8 model?
Wow this works absolutely amazing! I'm impressed how much freedom you have for all kind of scenes and angles and it still works very good. Can't wait for your next model! Maybe Cumshot? This never quite worked as an still image. But a video would be amazing!
I heard 24GB of vram is the recommended for training loras. I wanted to try with only 16gb on my 4070. Is it worth trying or no?
I can't tell you, I have 24gb. the settings that I've mentioned in the description maxed the vram out. It should be possible if you only train on still images or reduce resolution and frame count. good luck!
I have a 3060 12gb and I'mtraining loras, following this article: https://civitai.com/articles/10335/hunyuan-video-lora-trainning-with-gui-in-windows
Should train using Sladkislivki as a reference.
Now we only need cunnilingus. That would be awesome
How is the LORA, with diffusion pipe or other? formed, what are the necessary PC capacities?
nice