WAN 2.2 14B Workflow
A workflow for WAN 2.2 14B with multiple improvements for speed, flexibility, and usability:
Auto resolution calculator – automatically adjusts resolution based on a maximum dimension.
Video length in seconds – define duration in seconds instead of frames.
Fast interrupt – integrates BlehModelPatchFastTerminate to stop WAN generations more quickly.
Workflow Options
The workflow includes several features you can enable or disable to customize the output:
Enable Sage Attention – Enables the use of Sage Attention during generation.
Enable Use Tiled VAE – Uses tiled VAE decoding for handling larger resolutions.
Enable Interpolate – Interpolates the video to 50 FPS for smoother motion.
Enable Upscale – Upscales the video using the model selected in the “upscale” group.
Enable Upscaled Video Show/Save – Displays and saves the upscaled video.
Description
Fixed error "SimpleMath.execute() got an unexpected keyword argument 'c'".
Added Upscale function.
Added Interpolate function.
Added option to use Sage Attention.
Added option to use Tiled VAE Decoder.
Now it uses Power Lora Loader from rgthree package to manage LORAs.
FAQ
Comments (48)
The manager o comfy ui cant detect what it needs. any links for the loras and everything?
Hi mrfuckinnice, here goes:
Diffusion Model (GGUF version)
Select a version that better suits your VRAM size then download the low and high noise models, I use the q4_k_s. Put the files in ComfyUI/models/unet
https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF/tree/main/HighNoise
https://huggingface.co/QuantStack/Wan2.2-I2V-A14B-GGUF/tree/main/LowNoise
LoRA
For lower steps count. Put the files in ComfyUI/models/loras
VAE
Put the file in ComfyUI/models/vae
Text Encoder
Put the file in ComfyUI/models/text_encoders
@JCB2 Beautiful! thank you so much. This really helps for beginner!
also, should i just let the upscaler scale at 1.00 or i can change that to 1.5?
@mrfuckinnice Leave it at 1.0. If you want a higher upscale change the upcale model instead. The upscale factor is in the name of the upscale model, like 2x or 4x
@JCB2 im using 3060 12gb. so ill get the q2 high and low. that s 5 + 5 gb = 10gb. Is that how it works? and also. if i turned off the high lora but turns on the low lora, does that mean my steps needs to be 10 high and 4 low?
@mrfuckinnice get the q4_k_s, each model is loaded separately. So each individual model should have less than 12gb. I didn't understand your second question well... keep both high and low loras enabled and with same stepcount on the ksampler. I also have an 3060 12gb and the defaults values of the workflow work well, so you can just use them as they are
@JCB2 ive been using q2 and the result are horrid. Im downloading q4 right now. Thank you
@mrfuckinnice anything bellow q4 gets too degraded to be useful
@JCB2 I didn't have the "unet" folder, so I created it, I updated but I still can't put the GGUF files on ComfyUi
@eatme try putting the .gguf files in 'ComfyUI/models/diffusion_models' and see if it work.
@JCB2 Yes ! It's good ! Thank you very much !
@JCB2 Is there a way to make long videos without them looping or repeating? 🎥🤔
@eatme You got some control over that by using a more detailed prompt and specifying the motion of the scene, like a camera movement, a person that appears on the scene or an interruption of the action. But control is limited and the model not always follow your prompt very well.
Another strategy is using the last frame of the video as the start frame for another video, and join them after (I did this on this video [NSFW] https://civitai.com/images/103072821), but this strategy also has its flaws, as the model doesn't know exactly what was happening on the last video and at which speed, the movement can get a little sloppy between videos.
@JCB2 Hi JCB2,
I’m an indie game developer. I was previously using the web UI, but after seeing your amazing Wan2.2 I2V results, I was totally blown away – they’re absolutely perfect for game assets and cutscenes!
I’ve just installed the official ComfyUI. My question is: If I download your workflow JSON and put all the files you listed above in the correct folders (High/Low Noise GGUF in unet, the two lightx2v 4-step LoRAs in loras, the VAE, and the text encoder), will everything just work out of the box?
(My GPU is an RTX 5060 Ti 16 GB)
Thank you so much for sharing all of this – your work is seriously inspiring!
@dackphoenix186 You'll need to install the ComfyUI Manager extension to detect and install the custom nodes used in the workflow. It's a very straightforward process, you can check a tutorial if needed. Then, if every custom node is installed and you downloaded the models to the correct folder, check if the paths and filenames are correct for each Lora/model on the workflow, it can change depending on your folder structure. After that you should be all set to go and start generating videos (should, but other problems related to ComfyUI misbehaving can arise...).
@JCB2 I'm actually having a lot of issues with downloading the custom nodes as ComfyUI-bleh so Ive been stuck not being able to install 3 custom nodes, tho not sure if its becuase i couldnt install this file first. Please help me. This has been a very frustrating experience as a new user of comfy. Some people have said cause I need the manager, but for some reason my manager is missing and i can't see it anywhere.
Hi @Moshindo, you need to manually install the custom ComfyUI manager. The one prebuilt with ComfyUI doesn't work very well. You can follow this tutorial on how to install it: https://youtu.be/Si6ShFRFd0o
Then you can use its menus to easily find and install missing nodes
@JCB2 Ah I see yea i figured it out. Also is there a way to find some of these LORAs and embeddings that you use? I'm looking at your workflows and i see some Loras and embeds you use tht are different than the ones you say we should install
@Moshindo I don't use embeds on my generated videos. The Loras I use are indicated at the Generation Data of the videos, just click on their name to go to the download page. I don't use any custom or special Lora, you can find them all here at Civitai
非常喜歡妳的編排與功能，看到自動計算長寬我都不知道工作流還能這樣玩，謝謝你 !
Are only GGUF models possible to use in this wf or can models like SmoothMix also be used?
You have to add a Load Checkpoint node and substitute the GGUF loader ones, connecting them to their respective high/low lora loaders.
@JCB2 Is it possible for you to add a start to last frame option or a looping option to this workflow?
@maaade I tried it once, but I couldn't get an implementation that gave me satisfactory results. Maybe on a future version I could try again.
@JCB2 Btw any tips to get such high detailed outputs like shown here? I have a 4090 and using q5_k_m but the outputs seem to look soft or not that highly detailed.
@maaade Are you using the lightning Lora? If you set your step count low and doesn't use lightning, usually the video looks low detailed and smooth
@JCB2 Yeah I am using the default settings in your workflow. Didn't change anything and added the lightning models you recommended. I input a 1080x1920 image and only changed the size slider to 960.
@maaade have you re-selected the lora on the powerlora node? Sometimes you have to do this to set the right path on your computer
@JCB2 Yeap, manually selected it. Also do you think switching to q6_k would help? Or enabling tiled vae and setting size to 1920? That possible with my gpu?
@maaade I use q4, so it isn't the quant. Your gpu isn't the problem also. Enabling tiled vae would only help if you are getting out of memory error when decoding. Are you using the smooth mix checkpoint? If so, try disabling the lightning on the powerlora, I think it comes already merged with the model, so you don't need to add separately. If this doesn't help, could you share a video with the "save_metadata" option enabled or export your workflow from the file->export menu?
@JCB2 No, I am not using smoothmix since I am using the gguf models. Okay, I will try and share it.
@JCB2 Oh I think it's because I didn't enable the upscaler and interpolation. Maybe that's why its looking kinda low res in my 4k monitor. Thanks for the help tho! But any tips to make the output follow my instructions? Like when I want a couple to keep kissing while the man massages her breast, the video makes them stop kissing by the middle. Even if I explicitly mention it in prompting.
@maaade I'm glad it worked! About the prompting, I think this is more of a Wan limitation, it doesn't always follow instructions exactly. You can try repeating the action you want multiple times in the prompt, or testing out different seeds until you find one that works. But using a LoRA for the specific action you want would probably work better, or even combining two LoRAs, like one for kissing and another for breast play.
@JCB2 Yeah I have breast play but sadly can't find one for kissing. Not for wan 2.2 :/
@JCB2 Yeah I have breast play but sadly no kissing lora is available for wan 2.2 :/
@maaade I found this one: https://civitai.com/models/1881060, but I haven't tried it yet. You can also use LoRAs for the Wan 2.1 14B on the 2.2 14B, they are compatible. Just add the same file to the high and low, and they should work
@JCB2 Oh thanks for the link! I also wish wan 2.5 makes it out to us. It has like insane prompt adherence and quality!
@maaade yeah, I hope they make it open like 2.2
@JCB2 Oh I also meant to ask, how to open the negative prompt or see it? I can only see the header but not the box to type it in.
@maaade you click on the grey circle on the header
every output i have is super blurry and noisy and im basically using the default settings for everything. I've tried to make edits to chage it and the video does seem to be generating and the prompt is being followed, but the video is just terrible quality and gets worse progressively from the beginning to the end of the generation. ive tried troubleshooting with ChatGPT but no avail. Any ideas @JCB2 on what i might be doing wrong?
This problem seems similar to what happens when you use a low step count. Are you using any speedup LoRA like Lightning or Lightx2v? Without them, you need to use a high step count like 30.
The default workflow settings should fine if you're using the Lightning LoRA.
@JCB2 ill give that a shot. i thought i was using it correctly but i did get them out of order and the low noise version has the weights screwed up and i accidentally applied it to like 50 generations. Hopefully that's all it was. I'll experiment, thanks again dude.
Hi, The 5B Wan2.2 workflow if working nice. But this 14B model is giving reconnecting error. It's starts low and increases ram usage to maximum instead of gpu usage. Default settings. Using 4070super 12gb, and 32gb ddr5.
Which model are you using? If you're using a quantized GGUF model, try using one with a lower q, like q4 or q5. This should help with ram issues
I am using a Q4 model. But your other workflow if working for me. So I will be using that. Thank you.