EDIT! (25.10.10): Okay, okay — I’ve added the HIGH NOISE model! :) Took me a small fortune to train it in proper high quality. For the workflow and more training steps, please check out my Patreon — link below. <3
https://www.patreon.com/polyhedron_ai
So, folks. My first LoRA for the WAN2.2 model (T2I, Low Noise) – mainly meant for single images(!), not short videos. The training set was originally built to reduce that smooth, unnatural (plastic) skin look – first for SD1.5, then SDXL, then FLUX, and now finally for WAN.
A High Noise model might come later; right now, training these LoRAs at all is pretty expensive … But anyway: here it is! <3 I’m sharing the “finished” version after 3,000 training steps, which is the one I personally use. If you feel it’s already a bit overtrained, you can grab earlier training steps on my Patreon, along with the comfyui workflow and a detailed guide on how to get it all running on a smaller setup. (I’m using a 3060 with 12GB VRAM.)
Long story short: give it a try. It makes things a bit nicer, though it’s not a total game-changer anymore for getting rid of the plastic-skin look; WAN2.2 is already super solid and refined … but it does sometimes have that “Asian-style” vibe, if you know what I mean, and this LoRA helps dial that back. Have fun!
Cheers from Germany, Frank
If you need more flexibility, please join my PATREON page, where you get more training steps and the final LoRAs and how to use it.
Description
V.1
FAQ
Comments (9)
- training these LoRAs at all is pretty expensive
then
- I'm using a 3060 with 12gb RAM
???
I've trained over 50 Wan 2.2 LoRAs on my 3060 12gb.
It takes a couple hours... doesn't cost me anything.
Okay ... How? Do you have a Tutorial for me?
@Polyhedron_AI - I posted this quite some time ago, but the basic structure still works just fine. It is a rough guide but it's all there and there's a lot of superfluous text. I have a video tutorial that isn't fully edited yet and when I finish with my current hardware tragedies I will likely tidy it up and post it. Meanwhile, if you have any questions about what I do just ask.
https://civitai.com/articles/18181/training-wan22-loras-in-windows-with-musubi-tuner-in-dual-mode-to-produce-one-lora-not-two-12gb
@leisure_suit_larry Thanks a lot. <3 I will look into it. :) Cheers! f.
@Polyhedron_AI Here is another training method that is really easy, you can use it with Z-Image Turbo and base, Wan2.1. Wan2.2, LTX2 etc.
ostris/ai-toolkit: The ultimate training toolkit for finetuning diffusion models
To make installation easier that github repo provides a link to a script that will easily install it for you:
Tavris1/AI-Toolkit-Easy-Install: One-click Portable Windows installation of 'AI-Toolkit by Ostris'
Here is a video from the person who made AI-Toolkit:
https://youtu.be/2d6A_l8c_x8?si=w5dly6j7ateqPgP6
Here is another video from someone else that explains things well:
https://youtu.be/zegMOOKy_6A?si=JbcKmwSKiqg0mRUS
On my first attempt, the results seem very promising! I'm currently using the WAN2.2-14B-Rapid-AllInOne, which has that distinctive plastic look even with different schedulers and samplers. By adding this LoRA, the issue seems to be fixed considerably!
what strenth you used? for both lora
I'm having incredible results with this Lora on Wan 2.2 T2V, the videos are MUCH more realistic! It works great with any Lora I use (NSFW or SFW). I'm using 0.6 strength on high noise and 0.3 on low noise, I'm getting much more facial variations and much more realistic skin texture and anatomy in general! You did a great job here!
Thank you. <3