this is a simple adaptation of RES4LYF (learn about it!) for the amazing SINGLE FILE!!! WAN 2.2 T2I modded model: https://civarchive.com/models/1927692 <--- GGUF version (can be easily changed for loading safetensors ...)
Update 1: credit goes to this Reddit commenter --> https://www.reddit.com/r/comfyui/comments/1n9d72v/comment/ncpg10f/ <-- he inspired me to test and create the only-LOW-WAN-T2I version too ;)
Update 2: for faster speed of image generation get this tweaked WF from the directly downloaded PNG images of this post: https://civarchive.com/posts/21941339
Update 3: for NSFW use --> https://huggingface.co/NSFW-API/NSFW-Wan-UMT5-XXL
P.S.: all other needed resources are on Huggingface <--- use the search ;)
Description
this WF uses only the LOW-noise part of WAN2.2: https://huggingface.co/QuantStack/Wan2.2-T2V-A14B-GGUF/tree/main/LowNoise
FAQ
Comments (24)
The Workflow get Amazing result! 👍❤️
<3
Why CFG 4.5?
you can use any from 1 and up to 8 (that is written in model description), 1 - is the fastest CFG, 4.5 - is my experimentally determined best for many cases, ... again - set your preferred settings ;)
Why only low noise? Is it to make faster with a low quality loss?
it was suggested in reddit to my original post about magic-wan-image WF = that low noise also works well with this node setup, so we can do faster generation with less memory consumption
@OliviaRossi using Wan2.2 Low Q4 with 4 steps and lightx2v at 3 of strength, it is giving very bad results
was able to fix it, it seems that I was using PyTorch attention and needed Sage Attention
@BigDannyPt do you need 3 as a strength? I'm working at 1 with lightx2v
@OliviaRossi no, 3 was for the high noise and 1.5 for low.
3 still works, but will always generate Asian people.
@BigDannyPt in my WF there is no high noise part!!! and if you need not Asians - just write "European woman" ;)
@OliviaRossi yes, it was a mistake from my side. And using only low Q4 I can get an image in 1 minute on my RX6800
@BigDannyPt great!
there's something strange.. if I set cfg in the first sampler to around 1-2 I get realistic images.. but face looks wrong ..bad teeth, funny face... if I raise the cfg above 2-3 I get nice face but the image becomes more... animated.. colors etc. look too smooth.
Any idea why?
I am using the A14B low noise Q8 t2v wan.. not the magic one.
I was playing a lot with this WF recently, and I've made slightly tuned version of it for Light + Low Wan 14b T2V, get it here: https://civitai.com/posts/21941339 <-- download the image as PNG and drag it on ComfyUI = there are tweaked settings inside for faster but nicer results
You know what's crazy? Using Krea nunchaku as a refiner for wan low noise gives similar (if not better) and faster results than using the High noise in T2I (not for video).
not crazy - but COOL = thanks for sharing this idea!!!
Mind sharing your WF? Would like to try this refiner thing.
I am new to ComfyUI, sorry.
I've loaded the workflow and some nodes were missing, the program installed them except ClownsharKSampler_Beta
I wasn't able to find it anywhere,,,
here it is: https://github.com/ClownsharkBatwing/RES4LYF
@OliviaRossi Grazie!
As always, the author is a rock star, continuing to prove that there can be an artful elegance to AI image generation. I'm just grateful that I already had all the RES4LYF nodes lol.
<3
I seem to be doing something wrong. I added in a lora loader. Do I use high or low loras? Also without loras; the second, final, output seems to be way to strong. Like setting cfg way to high on SDXL.
you need to tweak it for your setting (sampler + scheduler + CFG + lora)