Update Kijai's MMaudio custom node.

@KaptainSisay Thanks, that fixed it!

@SeoulSeeker спасибо огромное , у меня даже без подсказки работает!

Sounds like the video was not converted to 24fps for mmaudio

@SeoulSeeker maybe, but the model is bad at performing sex acts. Your workflow is good, and it has nothing to do with it😄

Same actually. Just soured my taste for couple of my favourite clips. Results are laughable and I was bound to listen to hot girls talking simglish for a while lol. The workflow is fine though.

If your video is not 24fps, use the comfy load and save nodes option to force 24fps, otherwise it'll be out of sync.

is really useful?

@SeoulSeeker I test it, I have a good result,only the shortcoming is the voice sounds too old

@SeoulSeeker Exactly, that's the easy fix, and putting the cfg at 8 reinforces the desired result.

i will try it.thanks

@SeoulSeeker add some more, you mean:"young voice" like this?

@zczcg use that and emphasize on it, and on the negative side put, " Old, Old woman," and the cfg on 8 and it should do the trick, also try the "fixed" seeds starting on 1 and so on, idk why but it helps a little more in some videos.

@SeoulSeeker Well then you leave me no choice! I have to try it out.

Thanks!

Hi, when you use 50fps in the input, what did you put in the force_rate? I am getting bad results with 32fps input.

Thanks

@DaAlb You need to force at 24fps or it won't work. I started exporting 5s segments at 24fps out of my video edits and using it that way only saving the mp3 result.

@tenstrip thanks a lot

@SeoulSeeker ahh ok. Thank you. Looking forward to trying this :)

@SeoulSeeker Oh wow, this is one of the few workflows that actually works without big problems. This is awesome! So first step works really well. Increasing the frames of the videos is really easy to get running with this Workflow!

The second function is almost as simple. There is one potential problem where comfyui can't find the mmaudio folder. Took me a while to figure that out, but that has to be added to the extra_models_config.json or extra_model_path.json. I think this is because I don't have comfyui on C:\ drive because I need more space for comfyui so I added to both files the line and after this and a restart it worked:
mmaudio: Y:\programs\ComfyUI\resources\ComfyUI\models\mmaudio

Most of the crazy all-in-one WFs that abound here are 90% bells and whistles, landmine bugs waiting to troll you. I'm guilty of this myself; it's not that automation isn't useful, it's just super intimidating the first time you wade in and nothing works. I promise you it gets easier. I know it's pedantic to say, but I didn't really understand comfy until I started building from scratch, or from basic skeletons. It really works. The core of all of them always boils down to a few key nodes. And you can dive into complex logic without tying your brain in knots over screwing up the python syntax. Or even knowing any python, for that matter. Also, as you build and test, you get an intuitive feel for the execution order, which was one of my biggest sources of confusion when I started. It's a great playground. Anyway, just wanted to encourage you to keep at it, it's such a great environment. Oh, if you haven't already- take a snapshot of your custom nodes, and make a backup of the whole comfy git folder (you can leave out models, input, output). When you inevitably need a clean install, having this prepared turns a nightmare weekend of forensics into a half-hour breeze.

@SeoulSeeker I did. But I got it to work! And its amazing! Do you know if there will be updates on this??

@SeoulSeeker Thank you so much for this!

@SeoulSeeker I clicked the download button and it downloaded as a txt file? Is that supposed to happen or am I missing something?

@SeoulSeeker oh sick, The more you know 🌈

@SeoulSeeker Hello, I tried your modified wf and it seems to be outputting to temp/MMaudio instead of output\MMAudio. Any way to fix this? (I'm also a noob)

@SeoulSeeker Thanks for the response! I meant to say that it's outputting into ComfyUI\temp\MMAudio instead of ComfyUI\output\MMAudio. It seems to be using temp as a base folder no matter what i put under prefix

@SeoulSeeker Oh well, a minor inconvenience I suppose. My guess is that the audio sampler makes a copy of the input video then muxes it with the audio file it generates, but doesn't cleanup the duplicated video. Anyway, thank you very much!

@SeoulSeeker Sometimes the audio is still too fast for some motions but I only save the audio, and when it does sync up it's pretty exact and the audio clip is also the same length as the 25fps clip that I export and run through it.

to use GPU with ffmpeg
ffmpeg -hwaccel cuda -hwaccel_output_format cuda -i input.mp4 \

-filter_complex "fps=30" \

-c:v h264_nvenc -preset p7 -cq 19 output.mp4

Even using cpu it works faster
2 minutes video takes few seconds

Thanks for information, I prefer 64 fps video, I don't need to do double interpolation coz

Recommended FPS for MMAudio

30 FPS — optimal

25 FPS — acceptable (European standard)

15 FPS — minimum, lower than this may cause poor lip-sync and motion analysis

60 FPS — also OK, but the model will internally downsample the framerate to ~20–30 FPS.

❗ Important

MMAudio is a multimodal audio-visual model, and it normalizes the video to its internal FPS, usually 20–30 FPS.

Therefore:

Converting from 60 → 30 FPS will not reduce quality for MMAudio.

@ValuedRender If you increase the framerate without interpolation then the video will just get sped up, that makes no sense.

@LastAssignment No it's not

Changing the FPS of a video with FFmpeg does NOT slow it down or speed it up by default.

When you re-encode a video and simply set a new FPS (for example, from 16 to 24 FPS), FFmpeg keeps the same video duration. It does this by adding or removing frames so the playback speed stays exactly the same.

If you increase FPS → FFmpeg duplicates or interpolates frames to reach the new frame rate.
Speed stays the same.

If you decrease FPS → FFmpeg removes some frames.
Speed stays the same.

The only time FPS changes the speed is when you apply special options like -vf "setpts=..." or -r on the input, which explicitly modify the timing.

So, simply changing FPS while re-encoding does not make the video faster or slower — it only changes how many frames it contains, not how fast it plays.

@LastAssignment I'm telling also about double interpolation in workflow example.
Do you really need to do it twice ?
I doubt I want to do interpolation for every generation at all. It's just wast of time.
Better option do interpolation for good generations, and then when you have 40 or 60 fps videos, convertation to lower fps takes seconds even on cpu.

@ValuedRender  Interpolation with the popular RIFE VFI node takes around 1 second on modern hardware (with this easy fix https://github.com/Fannovel16/ComfyUI-Frame-Interpolation/pull/102), it's a complete nonissue. Reencoding multiple times with FFMPEG just adds another step of quality loss. The best way is to just integrate everything to use the raw images/latents as much as possible in a single workflow to keep degredation as low as possible. You can disable steps inside the workflow if your only intent is to test stuff before upscaling/interpolation.

If you already have an upscaled/interpolated video and just want to add audio then the best way is to down/upscale it to 25FPS and only use that interpolated video for mmaudio and combine the original files.

Someone make a node to re-encode with ffmpeg inside comfyui, because I sure as hell wouldn't be doing that manually, way too lazy. Probably gonna vibecode one myself quick.

@LastAssignment > Reencoding multiple times with FFMPEG just adds another step of quality loss.
What multiple times ? This can be done ( decreasing ) at the end for "good" generations, without losses for original video. Imagin how fast ffmpeg will be on 5090 RTX if you know how to enable GPU support.
Just wast of time do interpolation for all generations. I'm done )

I got this too. Any fix?

@logmarch2948 same

solved:

**error was outdated nodes from nodes list tab, update nodes, double check the model name, if still persisting, rename the main mmaudio model to something else and test it again after making sure your relevent nodes are up to date

@SeoulSeeker If I want to work with longer clips (over 8 seconds), do you have any tips to keep the timbre consistent in the video?

@SeoulSeeker
May be using fixed seed for all chunks will help.

I'm trying clone of repo localy, with spliting incoming video by 8 seconds, processing and put it together.

@sushesousuo140801 This model phazei/NSFW_MMaudio produces very poor results. It makes anime sounds, noise, and so on. Everything is so bad that even short videos are unlistenable ( for me). Have to wait for better trainer )
Better option is to train on own data.

https://civitai.com/posts/25058188 25 secods

fixed seed

@SeoulSeeker I wish I had time for this ) 80% of success is dataset. pre-processing, noise removal this will take much more than 40 hours )

@LastAssignment They were and this is exactly the reason: i noticed it is the frame_load_cap which changes depending on mode and is less than the source frames. For a short clip I get 120 frames with AnimateDiff setting, 117 with WAN and IIRC 115 with LTXV. I guess there is a reason for it. But then again I created that clip with WAN in the first place, so why is the output not 117 frames in total and is 120 instead... maybe ffmpeg needs frames to be a multiple of a certain number and pads it on file creation?

@Shamino_The_Anarch I'm honestly confused as well :)