A GGUF conversion of this great work https://huggingface.co/Phr00t/WAN2.2-14B-Rapid-AllInOne
simple JSON workflow (you need only gguf loader custom node): https://civarchive.com/api/download/models/2099499?type=Training%20Data
umt5xxl and vae excluded. you can use separated ones from wan 2.1
Description
FAQ
Comments (37)
Thanks for nsfw v9.1 version. If someone wonders what's new with nsfw v9.2 vs nsfw v9.1 it's only different text encoder so with our gguf here thats non important because gguf don't have text encoder.
I think its important if the improvement was the text encoder then we need to know what it is...
@brahianvalles You can always change your encoder yourself for gguf. Search for 'nsfw UMT5' at huggingface.
@flo11ok874 why i always getting error when i use nsfw_wan_umt5-xxl_fp8_scaled.safetensor
@kumarkishank959811 Works fine for me. Did you change something in workflow? Only change should be old text encoder to new one.
@flo11ok874 thanks bro now its working good can you tell me a different nsfw_umt5 vs original in expand
@kumarkishank959811 You can ask author, I only know it sometimes helps with 'nsfw' - https://www.reddit.com/r/NSFW_API/comments/1mv4z9y/new_nsfw_wan_encoder_released_today_improves/
In my testing, I found that compared to V9, the V9.1 release introduces significantly larger variations in facial consistency. This issue makes the generated faces change too much between frames, and I’m not sure how to address or mitigate it.
v9.1 is nsfw version meaning added many nsfw Loras this is the reason of changes and possible worse quality.
v9 is no-nsfw version and don't have merged loras other then speed and quality.
For quality, and face consistency use v9 + optional 1-2 Loras you like.
Nsfw version is for easy 'fun stuff'.
@flo11ok874 Is it possible to use the v9 with the Wan2.2 General NSFW Model?
I've tried v9.1, the movement is great but the faces are just......uncanny in anime style, so when I saw your reply I turned to v9 + Wan2.2 General NSFW Model instead.
However it doesn't seem to generate proper result. The character is barely moving in a lot of seeds I tried, and even if they move it's nothing like the example videos. Just kinda confusing.
Edit: Nevermind, I made some stupid mistake in the workflow......It works fine.
Silly question maybe. If I want to use some nsfw loras, whitch one must be? 2.1? 2.2 High? 2.2 low?
2.1 and 2.2 low only
CivitAi working like crap again? lol
Great work. BASE is the best for my environment, it's amazingly steady and realistic. Somehow v9 generates very blurry and unatural shape shifting images. By any chance I'd like to try v2 - v7 GGUF...
"unnatural shape shifting images" That´s right. the created persons do some kinda "Terminator" Stuff. If told to turn around, the head and the front of the person twisted itself to font in a fracture of a second.
Yeah, like the "Terminator" once did. Turning from back to front, without rotating.
it changes the faces into a default caucasian look despite having Asian reference images
I don't know what model you are using, but I have never had this problem.
There is a big issue with the NSFW version. All the charcters/girls in the video move always up and down. This make the model totally unusable for anything.
+1
Even with no prompts, they do that.
I wonder also, if I say in the prompt "She turns around and shows her back", that "she" do that, but turns around to front inside herself in a fracture of a second. What I don´t want, in that Scene.
I'm aware of this and been working on new merges to improve this. I recently released v10 of my NSFW merges where it should be better, but can't say it is completely removed.
I was trying to use FLF2v (first and last Frame) with the provided workflow and rapidWAN but in a 5 second video, for example, the character stays 4 and half second doing her movement in frame1, then suddently she appears already in frame2.
The irony with the NSFW GGUF version is that not only do you have to use the VAE and Clip like any other normal version of WAN. But what's more, not all Loras for WAN 2.2 work. Only Loras for WAN 2.1 work perfectly.
So what's the point of using the Rapid WAN version, if you're still going to use the VAE and Clip...
And not all Loras for WAN 2.2 work?
Wait! What? For real? Thats wild ... Oh Fkn Outstanding if thats true! ... I just came here to find out what the deal was with Loras not working, being able to use 2.1 loras on 2.2 sounds like fun, but that would strongly suggest that this is actually a 2.1 creation not 2.2
"low noise" WAN 2.2 LORAs do work, especially with T2V where it is completely based off of the WAN 2.2 "low" model. I2V is a mix of SkyReels 2.1 and WAN 2.2, so both should have decent compatibility. I am the original author and did this for the NSFW merges of v10 Rapid AIO where both WAN 2.1 and 2.2 LORAs were used just fine.
@ronikush I'm not even sure what went wrong for me but it refuses to work for 2.2 while it's perfectly fine for 2.1.
I get a channel error when trying 2.2 VAE
@Phr00t_ no, for of course i tried to put only lowin low lora, don't work. I tried with many Wan 2.2 Loras, nothing, the result is just awful or no result at all. Yes, i only tested with T2V.
@NikeTakeuchi Wan 2.2 14B does NOT use the "Wan 2.2 VAE". Only the Wan 5B model uses the "Wan 2.2 VAE". Even the original official Wan 2.2 14B uses the Wan 2.1 VAE.
@Phr00t_ Oh I see, thanks alot man.
You are my hero, thank you very much
Keep up your good work, is the only model can I use to test WAN2.2 due to limitations on mi HW. Thanks.
Version v10 is out. Please share the latest gguf. Thank you very much
pleeeeease!
huggingface.co/befox/WAN2.2-14B-Rapid-AllInOne-GGUF/tree/main/v10
Can you guys do this for the 5b model please
I used all those models in the workflow that is provided in the description on this page. As a result, I receive animation of green frame. Why? What am I doing wrong?