This is the conversion to GGUF of the SmoothMix Wan 2.2 v2 - https://civarchive.com/models/1995784/smooth-mix-wan-22-i2vt2v-14b
Files available also in https://huggingface.co/BigDannyPt/WAN-2.2-SmoothMix-GGUF
The first two videos in each version, were made with the default Wan2.2 T2V workflow the others were by using the SmoothMix Workflow - https://civarchive.com/models/1847730
All were using umt5xxl-Q6_K GGUF and Wan 2.1 FP32 VAE
For the I2V, it is available from another user in hugginface - https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF
Description
FAQ
Comments (100)
Do you plan on doing I2V as well?
If he releases v2.0 for I2V, I might
@johnjohnjohnman there are i2v gguf's models at huggingface already.
@flo11ok874 You're right, don't know why I hadn't looked yet
@flo11ok874 can you share the link, that way I can add it to the model details so people are aware
What about Q1_S ?
I can do it, but in Q3 and Q2 it was already very degraded
okay
@fouchardmilcoupes311 do you want me to create it?
I can do it tomorrow
@BigDannyPt Honestly I was asking more for a joke, I didn't even think it was possible to go below Q2... But if you want to make it, i'll why not.
@fouchardmilcoupes311 you are correct, I thought I saw another option, but went into the quantize repo and Q2 is the lowest
@BigDannyPt If you're feeling motivated, you can try the wan2.2 remix. The creator has given his approval.👍🏻
@Mikeflower can you provide the link for it?
@Mikeflower OK, will have a look and try it and then I'll contact the owner saying that I did the GGUF conversion
@Mikeflower it is in now in hugginface - https://huggingface.co/BigDannyPt/Wan-2-2-Remix -, when I have time, I'll generate a video and upload into here.
more and more I wish for CivitAI to have an API way of uploading our models since the upload here is damn slow...
Smooth Mix Wan 2.2 GGUF what is this mean as compare to original ?
it means that it was converted to GGUF to take less space.
just a note, the normal one has some FP8 weights, it had to be converted to BF16 to be able to have the GGUF.
As comparison, the UNET part of the model is ~13GB while GGUF 18 is ~15GB, but GGUF Q6 is already ~12GB.
It is for people that don't have enough VRAM to have the original UNET in the GPU, while have a hit on the quality, it is possible to have the model in the VRAM to speed up the processing.
To add some supplementary explanation, if trained according to the standard conversion training process, the trained Q8_0 model will basically not lose quality. A large number of professionals have conducted comparative experiments on this.
@AI_Master_Workflow yeah, and I think that Q6 also has almost no loss in quality.
For this model, the space is not that effecient because there were fp8 weight that needed to be converted to bf16.
The bf16 of the unet was 23Gb
Is it possible to use the lighting lora to get faster generations ? I tried it and it didn't work. Is it baked with the model already ?
I think is already baked in, that's why it allows to use with 4 to 6 steps.
check the main model details
@BigDannyPt yes. The videos improved without any lightning loras. This is an improvement over base wan2.2. are there any smoothwan gguf models without lightning loras baked in?
About half of the time I get wierd confetti artifacts, and sometimes hieroglyphs that negative prompt can't remove. The clips that are successful look pretty good, though.
I don't think I saw any of that, but it might be due to the main model and what was baked in.
which models are you using ( GGUF version, text encoder, vae )
@BigDannyPt I'm using your Q8 model in the original text-to-image ComfyUI template. I switched to LCM sampler and that seemed to have fixed the artifact part, however, I still get symbols in circles on the sides of the screen sometimes, which is part of the training data, I assume.
@koto2091187 Usually Unipc/simple is pretty safe, or Euler/simple, also DPMpp SDE/simple
@koto2091187 In the orignal model's page it says that it should be used with euler a/normal or unipc/simple.
all my videos were with euler a/normal
@BigDannyPt okay, thanks for the explanation. I did try euler ancestral but the sampler was simple and beta, my bad. It's much better looking with normal, although it's not as good as Q8 base model visually, it's still very dynamic and clips turn out to be smooth as the title suggests.
@koto2091187 try euler not euler ancestral just in case
He did publish the I2V versions like a month ago i think (?)
https://civitai.com/models/1995784?modelVersionId=2260110
Will you be converting to GGUF?
That's for the old version, let me talk to him to see if he will be doing for version 2
@BigDannyPt Legend. I would do myself but I am in the midst of an epic internet outage (connecting through phone now) so downloading the safetensors is not really something i'd do atm :S
@etherloth Pastel says that he will be releasing a v2 for I2V, for the current one there is alreadya link for a gguf conversion from another user in hugginface in the my description
@BigDannyPt Roger, thanks champ
Благодарю чел! спасибо , давно ждал этого, + к карме))
You're welcome. I might start doing some more GGUF conversions in the future, at this moment waiting for the new I2V from Pastel to also convert it
@BigDannyPt можешь пожалуйста на него ссылку прислать. спасибо за ранее .
@gkir It is in this models description, the last link that is there, for the current I2V GGUF from another user
@BigDannyPt спасибо, нашёл, и прада много всякой мерзости в Т2V
@gkir ?
@qek может конечно мой переводчик плохо переводит. но иногда в место пенса . какой то лошадиный хуй))))))))))) или пенис на английском и есть лошадиный хуй ахахахаха
@gkir I don't think that any Wan model works with other languages besides English and Chines, or at least working correctly
@BigDannyPt тут не в модели дело текст энкодор за это отвечает , я пишу промты на русском в стандартный клип текст энкодор. . когда я использую смутчмикс модель одна без лорас генеталии рисует плохо, а с лорами гениталии животных
@gkir I would freak out to see a man with a big horse genital 😂
@BigDannyPt ну и правда часто встречались мерзость. так вот я разобрался! cfg увеличил от 2.0 и выше, в первом семплере, убрал лорас все! промт на английском. и пошло как по маслу. все работает , модель одна работает очень хорошо и что самое главное быстро, результат меня поразил . это t2v. так же в i2v cfg увеличил тоже ок!
I like this model and I'm using a i2v Workflow from DigitalPastel that works if I stick to official Wan 2.2 models but not with this BigDannyPT variant. I hooked LoadImage up to the start frame of WanImageToVideo but it's like it doesn't read it. It just creates based on the pos prompt. Anyone know how I can get this working?
This model is only for T2v, for the I2V version, check the descriptions that here is the link for the GGUF conversion made by another user in huggingface.
@BigDannyPt Thanks for replying man. I'd avoid a lot of headaches and wasted time if I would just slow down to read the descriptions better. There is just so much to learn with me image and video generation hobby. I don't quite understand why a t2v model won't take in an image. What the difference is there. I mean the difference in the making of it. But at least now I'm back on the right track.
@SmutSurfer1960 That is how Wan was build on. I only know about the 5B version that is a TI2V model, but is 5B, which means it will not have the quality of a the 14B
I’ve tried this version and the original and they are vastly different in terms of quality, the original looks very realistic and the gguf version looks very artificial. Considering that they are supposed to be basically the same why do they look so different. That being said they are both produce good results but in different areas. I wouldn’t say that they are the same checkpoint is this down to the workflow?
Did you used the same workflow?
Yes used smooth mix v2
@steve2167686 can you share the results?
also, which UMT and VAE did you used together with the GGUF?
Unfortunately, i have the exact same problem. The rendering is good, but not realistic. My gens look like semi-realistic pony images. I tried the exact recommended settings. Your gguf remix is realistic and excellent. Only smoothmix t2v is weird.
@Mikeflower it can be from the merge of the loras in SmoothMix, but I don't think I faced that issue, at least when using the normal template from ComfyUI, the workflow from SmoothMix would produce that type of fake skin
The link you provided, is that a link for the v2 I2v gguf for smooth wan or is that v1 gguf made by another user
What do you mean? There is i2v and t2v v2. The only missing gguf's model is t2v v1
there is no v2 I2V for now, DigitalPastel is still on the make of it, or at least, that what he told me
@BigDannyPt Hey man, can you convert the Smooth Mix T2V v 2.0 to ggufs? if not it's ok, the v1 is enough
@zeivon1999divinity613 this is the v2, I didnnot touched the v1
@BigDannyPt I see, amazing
Hi.why your file in huggung face in 1gb bigger than that file in civitai ?for example Q8_0.GGUF is 15.4 gb in hugging face but here is 14.3. exeuse me for curious
Not sure, but looking at the file here, it says to be 14.35 GB and in Hugginface is 15.4 GB, but when you press to download it appears a file of 14.3GB, the same as when you press download in CivitAI
Brother can you make gguf version of this i got 8 gb gpu . please its a request
提示词的如何写效果最好
I really can't tell, in that area I'm very bad, I normally go to an AI model, like ChatGPT or Gemini, and ask them to generate the prompt
I got the Q4 Km high noise and low noise. The generated video has nothing to do with the input image 🤣(it's also blurry,damaged and glitchy). dont think that's got anything to do with the models but more with me being a total noob
It's a t2v model. I was confused with the generations at first too. But the quality of q4 is pretty good.
May I ask how you do the conversion from safetensor to gguf? I haven't been able to find any guide on how to do it.
Look at the article that I've written. I explain how I did it
@BigDannyPt Thanks!
hi, may I know where the i2v workflow is? Can this model be used in an i2v workflow?
I used the normal one from comfyui templates.
But this model is for T2V, look at the description that I provide the link for the I2V GGUF or download the one from the models OP page
hi. where should i put umt5xxl-Q6_K GGUF. could you please explain which worlflow you use ?
I used the workflow from comfyui template.
GGUF can all go to the unet folder inside the ComfyI/checkpoints
Total noob here, what's the difference between HIGH and LOW? first I assumed it's noise, but why is the HIGH model more downloaded? is noise a good thing? 😅
Plus, is there some way to predict which model would work on my PC? at least some sort of 'spec range' ?
+) I've found out that it means HIGH =>'first half' and LOW => 'last half' of the diffusion process. So I have to download both the high AND low
@gjf34524351 yes, for Wan 2.2 you always need high and low or a all in one model.
Related to the one that works in your pc, best is to do the math of model + clip + vae + lora and see if they fit in your VRAM or at least VRAM + RAM.
If you are low on memory, you can also unload the model after the ksampler, that way you don't load the low while having the high in the memory
@BigDannyPt Thanks for the detailed instructions! about unloading the model(high) before proceeding to low, is there any workflow for this? I understand what you're saying, but I don't have a good example...
@gjf34524351 i normally use the VRAM Debug node from KJNodes, but you could also start comfyui with the --disable-smart-memory
@BigDannyPt You mean pass this parameter '--disable-smart-memory' when launching the main.py right? does this make it automatically work? Thanks!!!
and also(last but not least), does this tweaking make me able to use twice the parameter size model than the model I'm currently using?
@gjf34524351 yes, it is to set the parameter in the same command as the main.py.
I'm not understanding the second question. Are you saying that you want to use a model that is double the size that you are using right now? I don't know the model you are using right now, but you always need to have enough memory for the model + clip + vae + loras + whatever you are using the workflow.
@BigDannyPt Oh I totally forgot about your first reply 😅
I'll do the math and find out myself.
Thanks!!
Thanks for the work. Please specify more clearly in description or name if its I2V or T2V another time.
since in the description it cleary says that is for v2, and v2 is only for T2V, there is no need
It's so smooth, I feel like just leaving the framerate at 16fps.
Thank you, for some reason, this is smoother for my pc too for some way, thanks
where to get the Smooth Mix Wan 2.2 T2V v2.0 gguf models?
answer in the description of the model
Really nice model, runs flawless and fast and sticks good to the prompts. However, the results seem to be quite "plasticky", is there any advice to improve the skin and genital image realism? Any recommended setting or Lora?
This is yet another T2V model that doesn't know what a penis looks like...