It seems to work with this node, https://github.com/comfyanonymous/ComfyUI_bitsandbytes_NF4

@hiben40387 not working - ERROR: Could not detect model type of: C:\models\checkpoints\wan21NF4_i2v14B480pNf4.safetensors

@ai_wifus That is strange it worked for me. city96 just released the GGUF you can give those a try should be more easily compatible

https://huggingface.co/city96/Wan2.1-I2V-14B-480P-gguf

@hiben40387 This custom node is working, but when I run the workflow, I get an error saying some tensors are on the CPU while others are on the GPU. BTW my GPU has 12GB VRAM. Is there a way to fix this?

@ai_wifus Reading through that github it seems it's a memory leak error there is a someone made a fork to fix it that being said, for 12GB you are better off using the gguf Q4 from the link I gave, use it with these rwo nodes. These two together considerably cut down vram needed.

https://github.com/pollockjj/ComfyUI-MultiGPU

https://github.com/city96/ComfyUI-GGUF

@hiben40387 thanks for the reply! currently im using the q5 GGUF it works great!

@hiben40387 I have 12gb as well, ran into same issue can use the --gpu-only argument and I read about the other node fixing mem leak (hopefully other nodes also fix).
I have been testing full, vs gguf vs nf4. I assume nf4 is easier on memory than the full though right? GGUF is nice but slow and I found I can run full 480 models or 720 with 480 latent size. Or is GGUF and Nf4 about same speed wise?

I am still getting round to this test I assume it is a unet I know some nf4 like flux ended up being checkpoint loads

I had used this NF4 model with the node mentioned and I get nearly the same Vram usage(about 21-22G) and speed with official fp8 scaled model, I dont understand why but I guess I won't use this anymore.

I'm also getting the same error message. I'm using the Swarm UI, and I have an RTX 4090. I downloaded the 480p version. It's not working regardless of the frame rate. Can you provide a solution?

You get this error if you use the Comfy UNet loader node (which would be in existing workflows), you need to use the one called "Load FP4 or NF4 Quantized Checkpoint Mode" from https://github.com/silveroxides/ComfyUI_bnb_nf4_fp4_Loaders/blob/master/__init__.py#L178C16-L178C58

This custom node is working, but when I run the workflow, I get an error saying some tensors are on the CPU while others are on the GPU. BTW my GPU has 12GB VRAM. Is there a way to fix this?

https://blog.comfy.org/p/wan21-video-model-native-support

I just found example workflows: https://github.com/comfyanonymous/ComfyUI_examples/tree/master/wan

is the NF4 node available in the comfyui manager install custom nodes category in the comfyui manager?

@dims2 Just to check as I have been considering getting 720 even to make 480 videos, doe sit use more vram via text encoder for same size video, example teh sample 512 x 512 run both on 480 and 720 wil 720 eat more memory?

Custom node don't create new node, it replace exists node and rename it to "Load FP4 or NF4 Quantized Diffusion or UNET Model"

@mfireson ive followed the instructions about choosing channel : dev and then look "CheckpointLoaderNF4" and the file pops up in the "install custom nodes" when I hit install, it says that the it will not install because of the level of security parameters. so I am almost there but that node loader is not attainable. I will find a work around on installing that node because it is the final step i need to run this workflow. However, if you have info that is worth a shot, please feel free to share and lend your guidance. I was thinking of just changing the "CheckpointLoaderNF4" node to the diffusion/unet loader and see if that is viable as awork around.

as for now, I am stuck.

i fix the node but the rest of the workflow has so many bugs.

Here. https://github.com/kijai/ComfyUI-WanVideoWrapper

If a few nodes says they're missing, do these:

Pip uninstall torch -y

Pip uninstall torchvision -y

pip uninstall torchaudio -y

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

pip3 install torch==2.7.0.dev20250123+cu126 --index-url https://download.pytorch.org/whl/nightly/cu126

@Therma i think i found the issue, i was using an older comfy ui file vs the stand alone portable version. ill report back when i have tried this version.

Sage Attention is a method thats been super popular for video models for a few months, usually paired with teacache. Sacrifices a small amount of quality for a substantial speed up in generation.

@RuggedPineapple ah. these add ons make it difficult to utilize the workflow.

@Noob_ee I dont run this workflow, i just saw it scrolling through, but as general help I currently have 3 nodes that implement Sage Attention. One from Flux-Lightning, one from hunyuanvideowrapper and one from kjnodes that you have to turn on the see beta option for it to show up in your node picker. Also, sage requires Triton which is a royal pain in the ass to install on windows but a simple pip install command on linux, so if youre on windows the juice may not be worth the squeeze.

That said on 40xx series hardware taking the attention functions down to 8 bit with sage bumps generation speed 150% to 180% so its kinda magic

@RuggedPineapple right, i am not blessed running linux like others on here who seem to not mention what OS some of these images are using. the juice is not worth the squeeze, i was over here thinking that i was doing something wrong. i am running a 40series but shy of the 4090. ill just have to wait for a better optimization for windows.

@Noob_ee WSL 2

@voboyso what is that?

@Noob_ee Windows Subsystem for Linux. You can use SageAttention easily while still running Windows (you'll do the generations and all that in WSL (Ubuntu Linux)). Just search "WSL SageAttention" for more info, but ere's the video I followed: https://www.youtube.com/watch?v=ZBgfRlzZ7cw. I had several issues after that I needed to address; might go smoother for others though. The guy is a bit all over the place, but you'll be on your way to faster generations by following it (and possibly figuring out some other issues after, on your own).

Make it Make sense

I made an FP4 quantized version and it's still slower than a Q6 GGUF on my 5070 Ti. I dunno if there's anything special that needs to be updated to get the improved speed from the Blackwell FP4 optimizations.

Seems like it specifically needs to be NVFP4 precision to get the speed advantage. MIT recently released an SVDQuant method that supports NVFP4, currently just for Flux, but they're planning to add Wan2.1 support. I tested the flux workflow in ComfyUI with the FP4 model and 8 step lora and it's blazingly fast. I can generate a 1920x1200 image in less than 10 seconds, 1024x1024 in under four.

@thaddeusk Amazing! Thank you for the information, much appreciated!