To understand the workflow and ow it works watch the video
Description
FAQ
Comments (70)
No one wants long videos. They want 5 seconds. Experts agree, users are congenitally unable to focus on anything longer. And they don't even want those 5 seconds to loop, because that would require an extra click, also it would make it somewhat more watchable, which would be intolerable. Your efforts are wasted.
so far 7 people missed the joke.
@davcha 12 now haha people....
@davcha It's possible to get a joke and still not find it funny.
@PaperTyger Or possibly it's not a joke at all. Perhaps it's just an accurate description of a depressing reality. To follow your example and contribute nothing does seem more sensible. Cheers to you for your commitment to fighting spam.
@Ponder_Stibbons k.
Listen to this....what if someone wants to make a video for .....another website, or even personal use/project? Did I just blow your mind....yup, you're welcome. Now I'm going to block you, to never read another brain wroth comment from you.
focus on the bright sight , the problem for me is people making diferent nodes with the SAME FUCKING name ahhahah
35 dislike .... its a joke
this workflow run in windows , Can't import SageAttention: No module named 'triton' ,i am checking the windows version ,if anyone can help , thanks in advance
You need to install Triton, looks like you probably don't have it installed.
Install the woct0rdho's unofficial Windows build, so you save yourself the hassle of attempting to compile it. That should fix the issue. Make sure your virtual environment (venv) is active for ComfyUI, and then run "pip install triton-windows", and if it seems to go OK, try again.
also, chatgpt is really great at solving dependency issues. upload the workflow and the console log and it will give you the commands you need to run in your .venv to get everything installed correctly.
I managed to get SageAttention to work with this article: https://civitai.com/articles/12848/step-by-step-guide-series-comfyui-installing-sageattention-2
It took a while but I learned A LOT in the process.
Hope this helps.
@civitaidirectly068272 i did it with this article too. And it fucked up my comfyui when I installed python. Had to reinstall comfyui and setup all the shit in comfyui again. You won't believe how much I hate python now.
Reinstalling has also removed sageattention again... it's SO AWESOME!!!!!! FUCK PYTHON!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
Steps to Install Triton on ComfyUI Portable (Windows)
Follow these steps carefully (based on the 2025 updated README and community guidelines). Always use embedded Python to avoid conflicts.
Open CMD in the ComfyUI root folder (D:\ComfyUI, where python_embeded and ComfyUI are located).
Uninstall any previous Triton installations (if any):
.\python_embeded\python.exe -m pip uninstall triton triton-windows
Install Triton-Windows from PyPI (the easiest and recommended way in 2025):
.\python_embeded\python.exe -m pip install "triton-windows<3.6"
The "<3.6" limits to versions compatible with your current PyTorch (2.9.1+cu128 in log).
If it fails due to version, try without limiter: .\python_embeded\python.exe -m pip install triton-windows.
Hello. I haven't studied your workflow much, the default video is 19 seconds. The main drawback is that the color/contrast transition is visible at 10 seconds. How can I increase the video duration and get rid of color/contrast spikes?
gguf gguf gguf
Does this work by sticking together two 10 second generated videos or four 5 second generations? If so can it be time stamped so you can tell it what to do when the time comes?
That might sound dumb cuz am sort of a newb to this but where do I put high noise loras? there's only 1 block for multilora picking, so do I have to put both there or ... ?
I haven't figured that out either. I tought wan22 always uses both high and low noise loras in two steps. But then there are workflows out there using wan22 with only the low or high lora. I don't get it either. This high low noise lora stuff seems absurd the futher I get into ai.
I don't understand why you should put it online in such an unprepared state; do it carefully.
Leave comments, align, everything should be beautiful
This doesn't even work. Tells me I'm missing like 5 or 10 addons. I'm new to all this and its complicated. Been trying for a few hours.....
Whatever forget this.
If you cannot find the nodes online then maybe try something else - ai is not for you
Install ComfyUI Manager to your ComfyUI, then show the missing nodes, install them. Also, you need to learn to read the command line output, and patiently see what fails - then google, ask an AI for help, and fix things one by one, until it works.
comfyui desktop not installing sage attention and triton automaticly is the dumbest thing in the world. It's abysmal
@ODBsGoonerExtravaganza if you can not make an intelligent comment try somewhere else - commenting is not for you
@i_m_l_O this works up to the point where you have to install sageattention. Python is such trash, it's crazy. The problem isn't that people are too stupid for this, the problem is how awkward and truly shitty python setups are designed
How can you write that it doesn't work when you don't even know how to install addons? Do you cry to your car dealer when you can't fill gas in to your car?
@ODBsGoonerExtravaganza look everyone hes a master prestige coomer!
Hi, just for you to know, the "Clean VRAM Used" node is not effective if the output have no connection, the execution order will ignore it.
That means you're missing nodes, go to Comfyui Manager and click on "Install Missing Custom Nodes"
This is my problem. there is no misssing Node. i dont get it but i am a New in this
@eifso123462 this is a hard task for someone brand new. I would get good at short ones first.
that means the model is not installed / downloaded. you can manually do it by searching the model i2v.highnoise bla bla, download it and put it in the folder where that model goes (comfyui/models). you should then be able to click it and select it from the drop down menu.
@eifso123462 Then those are probably custom nodes ? I haven't tried this workflow yet, so I might be wrong, but the only reason nodes would be red is if they're not installed.
I tried running this on a 5090 pod but it runs out of memory every time, is there something im doing wrong?
At what resolution and steps or framerate etc? Lower all your settings
Also tried it on my 5090 with 80GB of system RAM and it still OOM. I wonder what settings and hardware he used to run this mess at all! 🤣 Thanks for sharing though. I guess it takes a while getting used to 😋
Is it possible to run this without sageattention and something instead of that?
Trust me, you want SageAttention. It's worth it. It's faster. Figure out how to install it.
@aurelius It's not ready for Blackwell yet
@aurelius I just got it installed today for the first time and it completely fucked up my comfyui, so I had to reinstall comfyui and now have no sageattention again. I fucking hate my life.
I assume the damn python install fucked up comfyui. This is hell on earth.... it's such a fucking pain. It's everything everyone hates about python and the way it works with it's stupid pip pull trash. Back in my day we installed a programm and the programm knew what to do. Nowadays it's a rotte nfucking trash puzzle for the insane.
@stewi0001 My body is ready for the black well. Fucking shit
@stewi0001 Yes, it is. Install SageAttention 2.2. You can install 3 too, but you need to compile it yourself. 2.2 you can install the ready wheel.
@bluenightlagoon Because it isn't designed for users but developers. Have ChatGPT or Gemini help you with the install. They'll walk you through it.
@aurelius Sorrz dude, I am a developer,and python is still absolute ass. Modern programming is trash. I dont have chatgpt and I hear the free chatgpt is worse than the paid one, I dont have gemini either
@aurelius So I talked to chatgpt and it agrees with me. Python sucks as well as the whole installation process for triton and sageattention, which it says is the number 1 worst install process.
WHAT THE ABSOLUTE FUCKING SHIT!!!!!!!!!!!!!!!!!!!!!!!11111111111 It says I dont have to use trash python and I only need to copy sageattention and files manually into the custom_nodes folder! Why did I go through that fucking abysmal tutorial then. I thought that tutorial was good. But if this is true what chatgpt is saying that tutorial is absolut stupid:
https://civitai.com/articles/12848/step-by-step-guide-series-comfyui-installing-sageattention-2
@aurelius I'm sorry. I used now chatgpt to install sageattention and it was completely different than all the tutorials out there. It's really just copy and paste and also only sageattention, nothing else. No cuda, no microsoft tools, no triton. It all comes with desktop comfyui. Sadly chatgpt crashed during the explanation so I hope the installation of comfyui complete now.
In anyway, those tutorials are a huge trap. Well I say that but I haven't tested yet what I did now. But still I wanted to say thank you and say sorry for being so rude. But you have to understand how much time people on this planet waste with problems like this. It's a lot. Python is far from ideal.
@aurelius Fuck me, this is just getting worse. Now suddenly ChatGPT tells me that it is impossible to install sageattention to desktop comfyui and portable comfyui. There is only one weird portable comfyui version with torch and external python which will fuck up my comfyui desktop installation.... this is insane. This is absolutely insane.... I running in circles here for days now. This is absolute madness.
PS: needless to mention that the previous copy and paste of sageattention didnt work...
@bluenightlagoon For me, with RTX 5060 Ti 16GB, I have the following:
Python 3.12.10
Triton 3.5.1
pytorch version: 2.9.0+cu128
SageAttention 2.2
CUDA 12.8
Go to woct0rdho Github and find sageattention-2.2.0+cu128torch2.9.0.post3-cp39-abi3-win_amd64.whl
Download it and pip install it. That's the combo that will work. When you actually run comfyUI, you need to ensure it WILL NOT UPDATE YOUR PYTORCH VERSION. You need a specific torch version for it to work.
Though that's for my setup. I dunno what your GPU is. This will work with an RTX 50XX series at least.
@aurelius even chatgpt cant get it working... after 8 hours I reinstalled the fucking venv folder like 20 times and it still doesnt work. Chatgpt always tells me, allright this time it will work 100%. And always torch is the fuCKING worng VERSION because python is such a trash heap of a language that isnt backwards compatible even in the slightest. This is why python is just trash. You can install the latest python (I didn't) but you need to install pyton older python because shit python has zero compability in itself. It has ZERO backwards compability, the python code OF TODAY is the TRASH OF TOMORROW. It's so bad... and it just doesnt work.
I think I did it. This tutorial helped: https://www.youtube.com/watch?v=Ms2gz6Cl6qo
What I learned:
1. Comfyui desktop bad. I expected this to be the best
2. these custom installs/"""""""PORTABLE"""""""" installs are stupid. Because they are absolutely not portable. Your PC is after the installation a glas card house that you better never touch (change) any setting or you might break your """"""""""""""""""""PORTABLE""""""""""""""""""""" install.
3. python is not good
4. feels like walking on egg shells, I'm surprised it worked because things turned out completely different than the tutorial told me, like installed torch and cuda versions are higher than the sageattention build. I feel like this will bite me in the ass soon. I did everything exactly like the tutorial, yet my cuda and torch versions are ALWAYS(!!!!!!!!!!!!!!!!!!!!!) different. I hate python so much
of course it STILL does not work... this is all such crap
So I finally got it working. I now got SageAttention working. Is it faster? Maybe? Not sure. Seems same speed as before. The result looks kinda shitty. But can be due to randomness. But now my old the old video workflows don't work anymore. The default templates for wan. Those run out of memory now and I think I will go back to desktop comfyui.
This sageattention comfyui is one big bag of bugs even after you get it working. It's shit. I think it broke my complete cuda installation because crystools doesnt show my video cards anymore.
Allright, conclusion: thank you for everyones help. It's not your fault I'm raging, I'm raging because python is just so trash. It's a ductape programming language we all have to suffer because of this.
I now experienced my setup with comfyui desktop AND comfyui portable with triton and sage attention. Is the hype real? All videos I generated with sageattention looked half bad with lots of jump cuts, too fast animation speeds, very artificial motions and plastic or 3d looking elements. Could be because of the workflow, but it was constistent. Rendering time for a normal res video of 6s was 45 minutes.
With comfy ui the results of my workflow are more mixed, some videos are better some are worse than what sageattention workflow created. Without sageattention (comfyui desktop) I had less artifical 3d looking elements in the videos and constistent logical motions and speed. The rendering duration for 10s video is 45minutes to 1h and 20 minutes with sound.
The stability of comfyui portable with sageattention and triton is abysmal. Comfyui flatout crashes and elements like gpu graphs just disappear forever in the ui. I never had that with comfyui desktop.
So I my verdict is, there is zero benefit to sageattention, in my experience it gives you worse results in more time. It breaks basic wan workflows which will crash your whole sageattention comfyui everytime and break your ui forever.
@bluenightlagoon Your install is likely completely messed up. SageAttention doesn't affect any workflows whatsoever when it works properly. You just "plug it in," and nothing changes except it outputs faster. 7 minutes to render a video suggests something is severely wrong with your setup, like you're out of VRAM and it's using RAM. You're either rendering with CPU because your CUDA driver failed, or swapping extremely hard. SageAttention doesn't make you use more VRAM at all; if anything, less.
Regarding using AI for advice, you MUST make sure it's actually using the internet. Otherwise it's hallucinating everything.
As for quality, it may change a bit, but it should be mostly impossible to notice.
Anyway, it's not the end of the world if you can't get it to work. It just means you lose like 30-40% in speed. Make sure to overclock your GPU as much as you can too, as that gives speed as well.
@bluenightlagoon With SageAttention 2.2, I get 10.55 seconds per iteration. Without, I get 15.5 seconds per iteration. So the speed increase is significant.
кодек 'utf8' не может декодировать байт 0xc8 в позиции 49: недопустимый байт продолжения ..что за ошибка кто знает ?
*unsurprisingly* doesn't work
Hey guys, help me to solve that issue pls:
WanVideoModelLoader
Can't import SageAttention: No module named 'sageattention'
Switch over to the portable version
Works fine after adding modules. Wasn't difficult at all to get it going.
Can't find Wan2_2-I2V-A14B-High_Q8_(lightning edition)V1.1.gguf anywhere
me neither
2.2 Is so strange. I've got year-old generations from 2.1 720 model, hell 480 even, when I was just getting started with WAN, that look a billion times better than most of my 2.2 stuff. To match those I have to use fp16 version of everything and run at least 720 resolution. Even going full model, undistilled, or tks MoE, it's like pulling teeth just to get to the old baseline. That being said, starting with an ultra-closeup is usually fine, but any distance... man. That's apropos of nothing in your wf... just reminded me how annoying 2.2 is. Oh, the question. What does it mean when everything is all red? I have pentium III where does image recoder go I don't know how to make pictures move like you do I need a plugin I have extra ram slots whered o they go? I asked google it told me to leave comment on model page for fast answers.
5080 16gb 64ram 9950x3d 9100 pro nvme
pytorch version: 2.10.0+cu130
Python version: 3.12.10
ComfyUI version: 0.14.1
ComfyUI frontend version: 1.39.14
sageattention 2.2.0
tested your workflow with i2v
model wan22RemixT2VI2V_i2vHighV21+low
all others like yours,
steps 6/4 and frames 81
it takes 480+ seconds. and unfortunately when i place prompt more than 400 tokens - get OOM.
this is fast for 30 sec video.
but currently i used of this one workflow -
WAN2.2-I2V-AutoPrompt-Story QWEN
for 20 seconds video with no limit for prompt. it takes 180 seconds.
i guess 30 sec will + 150 more, or time will be the same as yours we speak only about i2v, without qwen generating promt this will cost more +15 min.
anyway for my PC more stable my wf.
i guess this information will be useful for someone.