I got error: with nvfp4 model as well. I have Blackwell card RTX5060Ti and use the official ComfyUI workflow: Node threw an error during execution.

ComfyUI Error Report

Error Details

• Node ID: 10
• Node Type: SeedVR2VideoUpscaler
• Exception Type: RuntimeError
• Exception Message: RuntimeError: Error(s) in loading state_dict for NaDiT:
  size mismatch for blocks.0.attn.proj_qkv.vid.weight: copying a param with shape torch.Size([9216, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 3072]).
  size mismatch for blocks.0.attn.proj_qkv.txt.weight: copying a param with shape torch.Size([9216, 1536]) from checkpoint, the shape in current model is torch.Size([9216, 3072]).

the same

https://github.com/Comfy-Org/comfy-model-tools

The exact conversion code used to convert/quantize the model is linked on the Model Card page.

The PR has not been released yet, but the models posted are verified working with that PR head as of a few hours ago.

Is it working in ComfyUI Stable release yet?

想要早点使用到seedvr2的nvfp4版本模型的可以先用我的这个修改版本：
https://github.com/nmvjhd/ComfyUI-SeedVR2_VideoUpscaler
已提MR：https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler/pull/604，待合并，仅仅一丢丢修改

======================
Any who want to use the NVFP4 version of the SeedVR2 model earlier can try my modified version first：
https://github.com/nmvjhd/ComfyUI-SeedVR2_VideoUpscaler
MR is submitted and pending merge：https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler/pull/604，just a tiny tweak.