This is a simple workflow for the WAN2.2_T2V_14b and WAN2.2_I2V_14b models, perfect for beginners. Essentially, it’s just a basic workflow that I’ve only slightly modified.
Key features: it works with SageAttention (which roughly doubles generation speed) and is friendly to systems with limited VRAM — it can run even on 8 GB graphics cards. Generating a 5‑second clip at 480×720 resolution takes around 5 minutes, while a 2‑second clip takes about 2 minutes. Hopefully, someone will find this helpful.
When setting up your project, locate the EmptyHunyuanLatentVideo or WanImageToVideo node and adjust the length parameter: - Set it to 33 for a 2‑second video - Set it to 81 for a 5‑second video.
A great step-by-step guide on installing Triton and SageAttention for ComfyUI Portable can be found here. It’s a straightforward set of clear, easy-to-follow instructions.
There’s also an excellent beginner-friendly prompting guide from the Civitai team here.
If you’re looking for the LoRA Lightx2v, you can find it here.
Kijai’s models are available here.
Wishing you creativity, inspiration, and happy generating!
Description
FAQ
Comments (23)
edit: thank you for the updated version!
(now i just need to get sage attention running without breaking ComfiUi portable)the node pack "comfyui-dynamicprompts" is unmaintained for over 2 years and won´t install in comfiui.
would be cool if there was a workflow with modern nodes
Do you mean a simple prompt node or something else?
it seems to be the prompt node.
comfiui tells me the node "DPRandomGenerator" is missing and suggests to install ""comfyui-dynamicprompts".
timtimber np, i will update it tomorrow
Akalabeth thx!, looking forward to test it
timtimber updated
Just wanted to say this workflow is running great! 3080IT 12GB VRAM. Generation times are 5 minutes. Any ideas why after several runs, the generated video quickly fades to grey?
I’m glad the workflow was useful for you. As for your question: I’ve generated 500+ short videos and haven’t run into that issue. It usually happens when the VAE doesn’t match the model, but if everything works fine at first and only starts failing after several generations, the cause is likely elsewhere. It could be a memory leak, but we’d need to figure out what’s triggering it. Have you seen the same problem with other workflows? Does it persist with SageAttention disabled? Do you get any warnings in the terminal? There could be many different causes here.
@Akalabeth I think I've worked out my issues. Thank for your assistance.
@dreadfulpirate Glad to hear you figured it out! You're very welcome.
i got
import torch
ModuleNotFoundError: No module named 'torch'
when running the test script. i see that i need to install cuda, i see in the cuda toolkit page that cuda 1.3 released, what version should i install? i didn't install sage and that yet, i'm still installing triton but couldn't test it.
My PyTorch version is 2.8.0+cu128. Not sure if cuda130+ is supported.
Akalabeth hey thanks!
i found out that comfy updated to cu129. Maybe that's why it itsn't generating.
I downgraded to cu128, but still can't generate. I ended up using a venv and the comfy manual installation. I still can't generate, after the diffusion steps, in the:
Patching comfy attention to use sageattn
Patching torch settings: torch. backends. cuda.matmul. allow_fp16_accumulation = True
100%
Restoring initial comfy attention
Patching torch settings: torch.backends.cuda.matmul. allow_fp16_accumulation = False
after that line, the comfyui gets closed basically, and the web ui says that it got disconnected.
I don't know if its because i don't have enough ram, or there's something wrong in the components versions since this comfy update (was literally yesterday so maybe that's why)
Anyways, i wanted to ask you, how much ram do you have or the workflow uses?
I currently have a Mobile RTX 3070 with 16gb ram (lenovo legion 5 pro laptop)
@fusky I’m really sorry, but I’m almost certain that WAN 2.2 won’t run with just 16GB of RAM.
I have 64GB of RAM, but I think 32GB should be enough for successful generation with this workflow. However, 16GB is definitely not enough, especially considering how much memory is used by the system, browser, and background programs.
If possible, I’d recommend upgrading to 32GB of RAM on your laptop. WAN 2.2 is quite demanding, and if GPU only has 8GB of VRAM, a lot of load shifts to system memory.
@Akalabeth yeah you're completely right. I just tried generating on a fresh install comfy and the default workflow, and it crashes after completing the first set of steps. Looks like i need to upgrade to participate with wan.
Looks really good, and I really liked your examples.
Thanks for your help! have a great day
@fusky You’re welcome! Really sorry you couldn’t get WAN running on your system. Thanks for the kind words, and have a great day too!
@fusky Hey there, friend! I found a way for you to generate using the WAN 2.2 model on your laptop. Try this option — the description says it works even for those with just 8GB of VRAM and 16GB of RAM. This all-in-one merge includes WAN 2.2 and other WAN-like models, CLIP, and VAE in a single container.
@Akalabeth wow! interesting! i just could generate a video of 31 frames in 10 minutes using one of the default wan rapid workflows.
I'm wondering if its going to be possible to use sage attention and teacache here.
did you have any opportunity to test it with your workflow wan22 for everyone? I just don't fully understand how to connect the nodes with wan rapid
@fusky Just connect the nodes in this order: Load Checkpoint > Lora Loader (optional, if you're using LoRAs) > Patch SageAttention (auto) > Model Patch Torch Settings > ModelSamplingSD3 > KSampler.
It’s all pretty straightforward, so I don’t plan to create a separate workflow for RapidAIO.
And yes — I tested Rapid for image generation with Sage Attention, and it worked without any issues. As for teacache, I don't use it because it significantly worsens the final result.
Maybe a stupid question. Would there be any benefit to adding teacache to this workflow?
It’s a very questionable idea, since enabling teacache has a terrible impact on hand rendering quality, leading to more failed generations. The slight speed boost just isn’t worth the trade-off.
@Akalabeth Thanks again for your input.