are you trying to install from within comfyui or are you git cloning the version of NAG node he posted in the nodes used part of the changelog into the custom nodes folder, if trying to just get the extension manager to try install it try cloning it into custom nodes

@satsukiyami cloning it worked, thanks!

I also get a VAE Decode error. I'm trying to use T2V w/ GGUF and I keep getting " TypeError: 'GGUFModelPatcher' object is not subscriptable "

@v_boissonneault282 Yeah, I've tried it on three ComfyUI setups, Desktop, Portable and Stable Matrix and it just doesn't work, I even swapped out the decode with the tiled version for stability and still, nope, not too sure what's causing the issue but I'll keep looking!

@v_boissonneault282 I've found a possible solution, run the text to video workflow, something simple, then turn it off and run the first frame and last frame, again, keep it simple, then turn it off and then run the image to video, I think the problem is tied to the VAE Decode needing something and for some reason the other two activate the search and install, try it, because mine works now!

@VicObXer thanks I'll try that right away!

So running the other workflows didn't work for me so I tried running it with LTX2 GGUF and it seems to work now. I could not get SmoothMix GGUF to work.

@DigitalPastel

HI i tried it but it didn't get the image from the f2lf at all. it's just random. how did you make it work

Ah ok thx

I tried what you suggested but it didn't work for me for getting I2V going. Instead, deleting the 'Clean VRAM Used' node just before the VAE Decode node got around the issue for me.

@eterelmel its work for me, thx dude

go to /custom_nodes in terminal, then clone this repo:
> git clone https://github.com/ThanaritKanjanametawatAU/ComfyUI-NAG-Fix

@Cyche Allelujah!

me too! and i have no idea how to resolve it.

could raising cfg help?

me too

Managed to get it working by disabling the modern node boxes

I am getting the exact same problem...

I am getting the exact same problem...

I am getting the same problem and I have no idea how to fix it, I am very new to this.

@valfuze14390 Hello, I encountered the same problem. How did you solve it? I'm a newbie. Could you please tell me how to "disabling the modern node boxes"? Thank you very much

What GPU do you have?

Install this node package: https://github.com/ThanaritKanjanametawatAU/ComfyUI-NAG-Fix

worked fine for me

@Cyche Thank you!

if you are running windows, I would use the official comfyui application.

i have the same probleme

same problem here...

@Krankheit @marcelinorafik01419

Navigate to this path: ComfyUI/custom_nodes/comfyui-mmaudio/mmaudio/ext/bigvgan_v2/bigvgan.py

Next, change two lines 368/369:

proxies: Optional[Dict] = None,

resume_download: Optional[bool] = True,

at least this helped me.

The crash isn't in VAE Decode itself — it's in the RIFE frame interpolation node that runs right after. The access violation is almost certainly an OOM (out of memory) VRAM crash on your 8GB RTX 2070.

Quick fixes to try:

Disable or bypass the VFI/RIFE node entirely — see if the base generation works first.

If you need interpolation, set RIFE to use CPU instead of GPU in the node settings.

Lower resolution to 320/480 or reduce frame count.

Enable --lowvram or --novram in ComfyUI launch args.

Close everything else before generating.

Your hardware can run this, but it's tight. WAN 2.2 I2V + VAE + RIFE stacked is asking a lot of 8GB VRAM. Try disabling RIFE first — that's the most likely fix.

- by Claude

Same. Dug around for about 3hrs before giving up. Hoping to come back to a fix someday

@zxcasddffdsxcv thank you so much!!!!

@zxcasddffdsxcv Thank you so much! I tried researching it myself and asking AI, but I couldn't figure it out all afternoon. You're amazing!

@zxcasddffdsxcv Bro, you are my Messiah

@zxcasddffdsxcv 10000 thumbs up!

Same here, i don't understand the problem...

Have you add the Loras ?

Try using https://civitai.red/models/1995784/smooth-mix-wan-22-14b-i2vt2v?modelVersionId=2513182 and updated loras, fixed this for me

@noisy_star yes 2214bi2v_i2vV20 step=30 cfg=2.3

it's not getting the input of my image at all

@kenjanjan Same for me, there is an issue with the workflow somehow :/

Same here and I reinstalled "ComfyUI Windows portable" to be sure.

@noisy_star I think something changed in the nodes / Comfyui, cause when I used older version yesterday everything worked but now? Even older workflow on older model does not work :/

So i made a mistake, I downloaded the T2V version of low/high noise instead of I2V that's why it didn't took the input. @WwolfXX @noisy_star

Same mistake (names are confusing) I used T2V checkpoints.
Thanks @kenjanjan !

@noisy_star no probs bro!

@kenjanjan Could you tell me the exact name of the checkpoint I should download then? Cause im confused xD

@WwolfXX when you open up the sample outputs above. you can see the link of the checkpoint it used. it should redirect you to the model, then download the other variant.

check the models if they are merged with or without LightX2V loras, blur is usually a lighting lora or ksmapler settings issue.

@Origion_0101 I don't have any lighting loras on and I'm using SmoothMixWan2214BI2V_i2vV20Low (and high). Should I be using something else?

you probably need the light2xv loras:

- https://huggingface.co/lightx2v/Wan2.2-Distill-Loras/resolve/main/wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022.safetensors

- https://huggingface.co/Kijai/WanVideo_comfy/resolve/main/Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors

"This workflow has an issue. After installing all the nodes listed by the author, a VAE decode error occurs. I switched to the wan2.2 sampler and the error went away, but the results are not as good as the author's. The problem seems to be with the New Subgraph node — I suspect the author did not package this custom node properly."

Apperently you need to create folder yourself. like this C:\ComfyUI2\models\mmaudio

@Paddington27 tks
I asked chatgpt for help, and it was exactly as you said. I've now managed to create the animation. During installation, due to a version conflict, I had to use Stability Matrix and select ComfyUI v.019.3.

And I also had to follow chatgpt's instructions.

@hoabinhfsodfikj Yeah hahaha we dont understand AI so lets ask the AI how it wants it. xD

Have you enabled the Workflow at the top in the Workflow Selection?

Did you fix this? Becuase I have the same issue

@Piraterage23 issue persists

I have figured it out. Yuo need to use loras that linked in description, with this loras everything works fine.

@borod333 I tried with those before even making this comment, still blurry. Out of curiosity are you NVidia or AMD?

That's because of the absense of extension "ComfyUI-Adaptive Prompts".Install it and problem can be solved.

same here

I had that problem at first, too. I spent hours trying to figure it out and finally found the solution: Download the new version of mmaudio, or just copy the fixed bigvgan. py file into your mmaudio folder. It’s located under Custom Nodes -> mmaudio -> mmaudio -> ext -> bigvgan_v2. There’s already a bigvgan.py file there—just overwrite it with the new one, and it should work.

https://github.com/hkchengrex/MMAudio/blob/57e19aa1069d7d1b0273810ead8ca50290bf7823/mmaudio/ext/bigvgan_v2/bigvgan.py

@LFPerfect is this with the NSFW model?

I don't think it matters which MMAudio model you use.

Hello i have that issue ,i did action you mentioned and thanks but i have always MMAudio FeatureUtilsLoader

TypeError: BigVGAN._from_pretrained() missing 2 required keyword-only arguments: 'proxies' and 'resume_download'. And if i re copy en entire folder for bigvgan v2 comfuy lose completly the installation of mmaudio and impossible to install a bigvan version.

@RobJhonson Yeah, I had that problem with my new ComfyUI installation too, because I wanted a clean folder. I then shared the error with ChatGPT/Gemini, and it told me to add the following to “custom nodes->comfyui-mmaudio->_init_.py”:

import sys

import os

sys.path.append(os.path.dirname(os.path.abspath(__file__)))

from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS

Just open "_ini_.py" in your editor and add it like this, then save it, and it should work—it worked for me, anyway.

I always use an AI—usually Gemini—to troubleshoot.

@LFPerfect thanks =>  in _init_.py i have now :

from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS __all__ = ["NODE_CLASS_MAPPINGS", "NODE_DISPLAY_NAME_MAPPINGS"] import sys import os sys.path.append(os.path.dirname(os.path.abspath(__file__))) from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS

but i have always the same issue for the audio

i reloaded the workflow and i restarted completly comfuy

BigVGAN._from_pretrained() missing 2 required keyword-only arguments: 'proxies' and 'resume_download'

My version of comfuy version 0.27.1 ( the last version)

this is that init ? C:\Users\xxxx\ComfyUI-Installs\ComfyUI (2)\ComfyUI\custom_nodes\comfyui-mmaudio\mmaudio or C:\Users\XXXX\ComfyUI-Installs\ComfyUI (2)\ComfyUI\custom_nodes\comfyui-mmaudio ?

there is different version : GitHub - kijai/ComfyUI-MMAudio · GitHub

and GitHub - hkchengrex/MMAudio · GitHub

and if we don't install nodes by "nodes extension management" comfuy don't find the mmaudio installed

@RobJhonson Install mmaudio as usual via the ComfyUI Manager so that ComfyUI recognizes mmaudio.

Then copy the fixed version of bigvgan.py into the specified folder.

If you restart ComfyUI now, ComfyUI will no longer recognize mmaudio—at least that’s what happened to me.

That’s why I asked Gemini/ChatGPT what the problem was.

Gemini then told me that something was wrong with the paths or something like that—I don’t remember exactly what—and told me to do the following:

import sys

import os

sys.path.append(os.path.dirname(os.path.abspath(__file__)))

from .nodes import NODE_CLASS_MAPPINGS, NODE_DISPLAY_NAME_MAPPINGS

Enter it exactly as it’s written here into init.py. If there’s already something in the file—which will likely be the case—delete it and replace it with this, then don’t forget to click “Save.” I restarted comfyui afterward, and then it worked for me.

The issue is on the node MMAudio FeatureUtilsLoader ,What should these knots be pointing at? because bgvigan is not a safetensors

@RobJhonson Copy your error log into ChatGPT or Gemini.

@LFPerfect  Okay, I started the whole process over from scratch and it's working now—thanks again for your patience.

Now I just need to figure out the audio part and where exactly to place the prompts.

Okay, the sound works, but what is it used for? Does it generate only noises or ambient sounds? Because no matter what prompts I use, I always get sounds—not human voices or sounds a human could produce.Could someone shed some light on this for me?

I asked ChatGPT, and that is indeed what it told me; could you confirm that this sound model is limited to this specific use?Thanks

@RobJhonson No, the MMAudio node (and the underlying MMAudio architecture) is primarily designed for generating sound effects (SFX) and atmospheric ambient sounds, not for human speech.

Here is a detailed breakdown:

Focus on SFX: MMAudio has been trained to generate sounds such as the clinking of glass, the rustling of wind, footsteps, or mechanical sounds based on an input video or text.

Lack of voice modeling: The model does not understand phonemes or human speech structure. If you try to force it to generate speech, it will at best produce vague, inarticulate sounds or noise that may vaguely resemble human breathing or mumbling, but never intelligible speech.

Different architecture required: For human voices (speech synthesis/TTS or voice cloning), architectures such as XTTS, Bark, Piper, or ElevenLabs APIs are used. These are based on entirely different training data (speech transcriptions) and models (Transformers for text-to-speech).

Conclusion: If you need a human voice for your project, MMAudio is not the right tool. Instead, you should add a dedicated TTS node (e.g., for Bark or XTTS) to your ComfyUI workflow to generate the audio component separately.

No, the MMAudio node (and the underlying MMAudio architecture) is primarily designed for generating sound effects (SFX) and atmospheric ambient sounds, not for human speech.

Here is a detailed breakdown:

Focus on SFX: MMAudio has been trained to generate sounds such as the clinking of glass, the rustling of wind, footsteps, or mechanical sounds based on an input video or text.

Lack of voice modeling: The model does not understand phonemes or human speech structure. If you try to force it to generate speech, it will at best produce vague, inarticulate sounds or noise that may vaguely resemble human breathing or mumbling, but never intelligible speech.

Different architecture required: For human voices (speech synthesis/TTS or voice cloning), architectures such as XTTS, Bark, Piper, or ElevenLabs APIs are used. These are based on entirely different training data (speech transcriptions) and models (Transformers for text-to-speech).

Conclusion: If you need a human voice for your project, MMAudio is not the right tool. Instead, you should add a dedicated TTS node (e.g., for Bark or XTTS) to your ComfyUI workflow to generate the audio component separately.

@LFPerfect Okay, thanks for the detailed answer and the confirmation—either use this tool for ambient sounds or turn to other models and workflows that require LTX-based models.I completely understand the need to use multiple tools. But after generating the audio, you still have to handle lip-syncing.

LTX models can do this now, but they still lack a lot of expressiveness.

Check out this account: https://civitai.red/user/Acaserio

@RobJhonson Yeah, I've seen accounts like that before, and they look breathtaking. But I haven’t really looked into it yet, since I’m still in the process of finding my own visual style.

But after a few minutes of searching, I found this:

https://civitai.red/models/2498991/dasiwa-ltx23-workflows-or-i2v-or-flf2v-or-t2v-or-audio

Why don’t you give it a try? The sample videos look really good with the synchronization, too.

I’ll definitely test it out when I’m ready.

@LFPerfect I’ve already tried an LTX workflow with audio; it worked, but it resulted in a singing voice (haha), even though nothing of the sort was specified in the audio prompt.

In my humble opinion, the solution is to clearly specify the action in the base video prompt for both MMAUDIO and LTX.

I ran a test using WAN video and MMAUDIO without a prompt for the MMAUDIO node, and it generated audio that actually matched the "base" prompt better.

I’ll be running more audio tests.

Of course, as you pointed out, the LTX process and audio results are more refined.

And the overall process is much more resource-intensive than WAN's.

I was having the same issue but installing the Lightx2v Loras mentioned early in the post fixed the issue

@Darsovin thanks.i loaded 2 loras lightx2v mentioned in Power Lora Loader (rgthree) nodes ,there are enable.But i have the same problem i have always the last frame in png ,so no video :(

@RobJhonson try anothor checkpoint with lightning included like this: https://civitai.red/models/2003153/wan22-remix-t2vandi2v

@LFPerfect thanks i will try also that, but for the moment is select Wan2_1_VAE_fp32 and i have a VIDEO in OUTPUT.

I had the same issue with the NAG node. I cannot get it to create anything more than a brown square no matter what length or size I make it though. And I can't get it to recognize the Clean VRAM as a latent output to save my life. Have you had a similar issue?

@Sweetpotatopie Nope, I didn't have that issue. But glad to see you figured it out! 😉

ok i just select "resize" In the node resizeimageV2 and the video is cropped to the scale of original picture that loaded in input

Got it, need to delete ImageFromBatch node and first frame will be yours making merged videos smooth again. I don't know what's the purpose of it besides ruining smoothness of videos longer than 5s, but without it everything fine. Workflow still not exellent, but very very good job overall!