LORD WINNOUGAN PROUDLY PRESENTS: WANNOUGAN 2.2 TURBO + NSFW
What is WanNougan 2.2 Turbo + NSFW?
I was asked to bring Wan2.2 to turbo and bake in an NSFW LoRA. Most finetunes only give you FP8, so converting them to INT8 Convrot is a no-go! If you want to convert a model, you'll need it in BF16/FP16 format.
Included are the INT8 and INT4 text encoders - pick one. Personally the INT4 works great.
Settings:
Use Turbo Settings in ComfyUI and it works best with Flashattention 2, not Sageattention 2/3.
4 steps, CFG 1
Credit Where Credit is Due:
This finetune includes the DR34ML4Y - All-In-One NSFW by c0ur4ge
Converted with Confidence: The Starnodes Factor
All of my models are converted with Starnodes' Converter!
Thank you Starnodes!
Included is the INT8 Convrot High Noise and Low Noise diffuser models. I have also encoded the Text Encoder to INT8 Convrot and INT4 Convrot to save you time and precious disc space.
👇🏻👇🏾Support me below 👇🏿👇🏼
💚 Workflow: (https://pastebin.com/nHWpSccV)
💙 Support on Ko-fi: (https://ko-fi.com/winnougan)
I'm saving toward a hardware upgrade so I can expand into Ideogram 4 and LTX-2.3 video LoRAs and deliver releases faster. Every coffee helps make it happen. ☕
🖤 YouTube — Tutorials and reviews: https://www.youtube.com/@Winnougan
💀 Discord — Releases, requests, and conversation: https://discord.gg/CJv5wceJaN
Description
FAQ
Comments (14)
OK few ask.
I2V or T2V?
What are the benefits of int8 convrot WAN2.2?
How are you happy with your model, will you be planning on some changes?
Either way I am happy that WAN 2.2 getting some love. :)
It doesn't work. The image is full of artifacts. And the link to the workflow in the description leads to the Krea2 workflow.
It does work. You can see that it works by the examples given. The issue is your workflow and inability to understand how to modify things so that it works. Instead of trying to fix it you just come here to complain. Nothing is plug n play with video/img generation, you'll often need to make adjustments on your end because no two installations are alike.
@DaddyWolfgang the workflow he linked is incorrect. Downloading models and workflows, installing nodes is not hard, but it does get hard when people release things without proper instructions and just assume everyone will know.
Downloaded the files, renamed them since they don't seem to mirror the actual filenames listed (filesizes are right though) and got the workflow from the .mp4, but I"m getting key errors ( 1074, in loadquantized_module
qconfig = QUANT_ALGOS[module.quant_format]
~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^
KeyError: 'int8_tensorwise' ), Yep I'm running FlashAttn 2 instead of Sage, and yep I have comfyui_kitchen in an otherwise working and completely up to date comfyui install....
Something is amiss
That key error seems to be for the Load Clip node.
I like it, very strong prompt adherence BUT the camera is quite shake-y, if you can fix that it will be top notch :)
And what is the workflow for this checkpoint?
making int8 versions of the wan2.2 custom models found here is possible. the process is a tiny bit involved and its not perfect, but the quality is acceptable IMHO...
please feel free to try the models i uploaded to HF. There are three to choose from
https://huggingface.co/Dewprisim/WAN2.2-14B-I2V-INT8
Krea wf for WAN???? WTF in the link
how you tried suicide? if not, maybe consider.
can someone please a share workflow for comfyui
nvm..... if you save the mp4 from the cover ..... that has the metadata for comfyui
15s clips in two passes, using a 5070ti paired with 9800x3d/96gb system ram:
- took 167s everything optimized with sage attention
- now takes 153s to produce an arguably better output!
I wish others would produce INT8 versions of their models as well, or maybe I should learn about doing it on my own!