Several people have asked about the workflow that I use. Credit to ZXTOPOWER for the workflow embedded in the excellent Korean Women Busansuji Lora. I have used it with minor tweaks.
The workflow generates 5s videos in about 80s on a 5090 RTX
Description
FAQ
Comments (39)
Thanks for this. A big fan of your work!
Your stuff is really top quality. These girls in crotchless panties would make a collection I am working on. Hopefully not to hard for you
Finally. A world where big tits always flop out.
Even on a 4070ti, still make 5 sec (80 frames) in about 55-60 sec. Then just fix the seed and upscale/interpolate with whatever method you prefer (the one in the workflow is fine too).
That seems super fast, at what resolution?
Can you share your configuration? I also have a 4070TI but the video memory is maxed out
@Light2020462 For 60 seconds? 284 wide x 384 tall... but that's mainly to check if a prompt works well or not. And 16 FPS. Then you can pick your favorite of the batch to either re-run with a fixed seed OR upload into an upscaler/interpolation workflow.
@a8631916328 I'm using the high/low Wan2.2-T2V-A14B-HighNoise-Q4_K_M.gguf as the Q4_K seems to be within the memory ok. Hope that helps. Like the other reply, start with a low resolution and make sure the prompt and settings produce something usable... then work on upping the base resolution or upscaling with an alternate workflow.
@sjohn567 Ah ok cool, yeah that makes sense, yeah I should try doing more tests at low res to save some time.
@sjohn567 Hey, im also using Q4KM GGUF, generating at 208x384 still takes me more than 100 seconds and going even as high as 360p will make it stuck. may i know if you used the exact same workflow provided or did you tweak some stuff?
RTX 3080 Ti (12GB)
32GB DDR4
@Vedamin Hey mate, I also have a 3070TI and really new to all this - how did you get yours to not get stuck? I'm getting stuck at 50% on the k sampler and assume i need to lower the res, not sure how. As for upping the resolution - I use TOPAZ AI to improve quality of videos
I have a 4070 Super with 12GB. I tried the sample NVIDIA girl prompt, left the resolution that is set by default from this workflow, using Q3_K_M. Although I only generated 3 secs, it took 148.75 secs. Removed SageAttn since I'm on Pop OS.
This is one of the fastest workflows I have ever used for WAN2.2. Thank you!
This worked great!
@sAInd_man I've been having problems with comfyui_layerstyle, which breaks both tokenizer and transformer, two Python libraries. I had to reformat my PC and now I'm going to install comfyui on a venv. The best version of Python for comfyui is supposedly 3.12, but I've read somewhere that 3.11 might work better for Wan 2.2. How are you doing?
Mine keep hanging during "patching comfy attention to use sageattn", any idea why? TQ
Can someone send me the links to the models and loras used in this workflow?
A good strat is to screenshot the Lora section of the workflow and ask GPT to find them for you. Worked like a charm for me.
Not able to find wan 2.2 T2V a14B high/low noise Q4 K S.gguf. I am new to this help me out
Download bot models from huggingface: https://huggingface.co/QuantStack/Wan2.2-T2V-A14B-GGUF and put it into the unet folder inside the model folder.
Were you able to figure it out? @theupsider's instructions were solid :)
Whenever I have used this it has come out really blurry? any ideas as to why? thanks in advance xx
Hello! I found out the reason it comes out blurry is that you probably have the workflow's LORA's active without having those exact models in your model/lora folder. As soon as I went through and added those loras to my workflow, it came out peerrrrty
My 5090 also achieved the same result, and it took the same amount of time, you are amazing!！！！
The video output from running this workflow on my RTX 5090 is blurry and unclear. Does anyone know why?
You save the world
You save the future
You are our hero!!!
Been trying to get this running, but comfyui keeps erroring out saying I don't have triton installed even though I've downloaded triton-windows and tested that it's working (by running their test-triton.py). I'm completely new to this, so any help would be greatly appreciated.
My workflow, Yet Another Workflow, was inspired by this one. This is really useful for learning.
Spam this error, I cannot open comfy ui after downloading mods for your workflow !
, ImportError: tokenizers>=0.22.0,<=0.23.0 is required for a normal functioning of this module, but found tokenizers==0.21.4.
Trying to install this was a nightmare, since I had to go down a rabbit hole to get all the dependencies I needed- but it works flawlessly! Thank you so much!
Do you have a I2V version? Your T2V is excellent!
yay! got my setup running and creating video via Text2Image! But had a problem with sageattention, can't pick any except disable. CUDA is 13.0 so I had to disable it, but it works.. NVIDIA gtx4070 12GB.
I guess I can't down rev CUDA. Any ideas? At least video is about 5.5 minutes to create, so similar to I2V for me.
jsut get an error on both
ModelingSampleSD3
TypeError: unsupported operand type(s) for *: 'NoneType' and 'Tensor'
i disabled both ModelingSampleSD3 nodes and it worked
how to increase the length? its stuck on 81
never mind. i figure it out...this is really a good workflow. very fast on my 3060 12g...any plans on a I2V version?
@emotionaldreams2 What's the solution?
@challasvarias01106 delete the "Math Expression" and "Length" nodes. then you can change the duration in the "EmptyHunyuanVideo Latent" node
i got it to work but i get a black output, i don't know what has gone wrong i cant quite fix it
Took a while to set everything up but all the generations I get are nothing but snowy static and only 2 were barely visible. Any tips to get it to work? I also can't seem to find the TorpedoTits_wan22_e20_high lora on the site anymore. Could you upload it to google drive so we can use it? Thanks!