Load CheckPoint DragNUWA

Every DragNUWA workflow in this pack starts here. This is the node that loads the actual video-generation weights into memory and hands you the opaque DragNUWA model object that every other node in the pack wants. Nothing works without it, and it's also where most first-time installs go sideways, because the weights aren't the ComfyUI-managed kind you're used to.

DragNUWA is a Microsoft Research project (arXiv 2308.08089) that open-sourced in January 2024. It's Stable Video Diffusion under the hood with a trajectory-conditioning stack bolted on: you give it a starting frame plus a set of drag points or a motion brush, and it hallucinates a 14-frame video where things move the way you pointed. Back in early 2024 that felt like magic; the demo videos still hold up as a look even though newer image-to-video models have largely moved on.

What it loads and what it asks you for

The inputs are the ones that matter, and there are only three:

• ckpt_name - a dropdown of files in ComfyUI/models/checkpoints/. The default is drag_nuwa_svd.pth, which is the name you should actually use. See install below.
• dimension - 576x320, 512x512, or 320x576. This sets the output resolution for the whole pipeline, not just this node. Pick once and keep every resolution-aware node in the workflow consistent with it, or the brush math quietly assumes one size while your image is another.
• model_length - default 14, and the model is trained for 14 frames. Leave it. You can change it, but every motion brush in the pack is generated against a model_length, so changing it in one place and not the other is a recipe for tensors that don't line up.

The output is a single model wire of the custom DragNUWA type. It feeds directly into DragNUWA Run, DragNUWA Run MotionBrush, and the brush loaders.

Installing the weights (the part people fumble)

The node itself comes from this pack, but the model file is a manual download - ComfyUI Manager won't fetch it for you. The README gives two options:

1. The original drag_nuwa_svd.pth (roughly 9.5 GB) off Google Drive - the official route, but Google Drive throttles big downloads without an account.
2. The much saner choice: the fp16 pruned dragnuwa-svd-pruned.fp16.safetensors from benjamin-paine/dragnuwa-pruned-safetensors, about half the size and visually indistinguishable. It's the port from the Enfugue project, and it's what I'd download first.

Either way it goes in ComfyUI/models/checkpoints/, and if you renamed it, rename it to drag_nuwa_svd.pth so the dropdown default matches. Chinese users get an official hf-mirror.com wget route in the README because HuggingFace direct downloads are blocked there.

One honest gotcha

The loader hardcodes cuda:0 when it builds the model. No fallback to CPU, no device picker, no AMD path. If you're on a multi-GPU box where ComfyUI isn't on device 0, or you're trying to run this without an Nvidia GPU, this node will fail in a way that has nothing to do with your prompt. It's an old port and it shows.

Also budget for the dependency list: requirements.txt is the full SVD-era stack - pytorch-lightning, open-clip, kornia pinned to 0.6.9, opencv-python pinned to 4.6.0.66, even streamlit and wandb. That's a lot of stuff sharing your Python environment, and it's the most likely source of conflicts if ComfyUI starts breaking after you install this pack.

Expect a long load. The first inference after loading also has to compile a few things, so a 2-3 minute pause before your first frames appear is normal.