I've just rendered a 28 second S2V generation using the Q4 GGUF, a quantised Clip and quantised audio encoder, and I have to say, even GGUFs aren't cutting it here...

Running on a 5070 with 12gb VRAM and 64GB RAM

The lip sync was completely off. The character was supposed to be singing but their mouth barely moved, and mostly moved at the wrong times making the video absolutely useless...

I think S2V, moreso perhaps than any other WAN model, is one that you either run on a supercomputer with the full fat models, or you just have to accept defeat with. There doesn't seem to be an option for us regular low-power consumer card users when it comes to S2V sadly...

Thanks for this. I was able to see the truth in it easy enough last night by swapping the KJ scaled fp8 high and low WAN models with, first, Q5K and getting immediately better prompt adherence when re-running the seed on a previously generated video. I had already stopped using anything but fp16 on most clips and text encoders, and use MultiGPU to push them to CPU anyway, where I've got plenty of RAM. Q8 is far and away doing things more to the prompt and with what looks like higher visual quality as well on my 5090.

Same

i checked out the youtube comments:

"RES4LYF Node is the issue. Disable it and restart the ComfyUI".

That got my past that issue.

I couldn't find the RES4LYF node, but disconnecting the get scheduler from ksampler and S2V extend made it work.

@abeslu425 I've just had this exact issue, the RES4LYF was in my custom nodes folder so I simply moved it to disabled and I'm back up and running - 2 bloody days I've spent on this :o

Did any of you get this error

AudioEncoderEncode

Input type (float) and bias type (struct c10::Half) should be the same

I'm stuck again now

Thanks for any advice :)

Could you be more specific?

ok guys, so after reading other comments and the info on YouTube, I found the issue:

I have node installed called RES4LYF, even though this workflow does not use it, you need to disable it, so I just manually moved it to the disabled folder and then closed and restarted Comfy and now this workflow is working great again.
Another workflow I use relies on RES4LYF so I will write a simple bat file that moves it to and from the disabled folder until the author fixes it.

I hope this info helps others.

oh crap, now I have this error

AudioEncoderEncode

Input type (float) and bias type (struct c10::Half) should be the same

cany anyone help please?

I've found this, which might fix it, but I don't understand it, hoping it helps someone to come up with a solution :)

Fix fp16 audio encoder models by rattus128 · Pull Request #12811 · Comfy-Org/ComfyUI

It's amazing what you can do when left to your own devices isnt it :)

I've fixed this too:
on your comfyUI startup bat look for this
python main.py

and add this to the end:
--disable-dynamic-vram

Mine looks like this:
python main.py --use-sage-attention --listen --auto-launch --disable-dynamic-vram

Again, I hope this helps some of you :)

After many updates, I need to add this line again, but now I need to add to line 247 of file called
Start-ComfyUI.ps1
in the scripts folder:
$pythonArgs = @('main.py', '--use-sage-attention', '--auto-launch', '--disable-dynamic-vram') + $networkArgs + $vramArgs