STAR Model Loader

STAR is a diffusion-based video super-resolution model (ICCV 2025, from Nanjing University's PCALab) that turns low-res, compressed footage into something you'd actually show someone. And this node is its front door: one tiny node that assembles the whole stack - the UNet+ControlNet checkpoint, an OpenCLIP text encoder, and a temporal VAE - and hands you a single STAR_MODEL bundle ready for the sampler.

Don't let the simplicity fool you. The first time you run it, it pulls down a serious chunk of weights: the light_deg.pt or heavy_deg.pt checkpoint lands in ComfyUI/models/star/ from HuggingFace, and the OpenCLIP encoder and SVD temporal VAE download through their own libraries on first load. No API key, no license gate - it just quietly downloads while you watch the console scroll. The model cards are MIT, so you're not renting anything.

Which checkpoint do you pick? light_deg.pt for web video with mild compression; heavy_deg.pt for genuinely mangled sources - cable TV captures, heavy bitrate murder, the stuff the community's "heavily degraded" fine-tune was made for. When in doubt, start light; heavy tends to over-restore clean footage.

The other three inputs matter a lot more than people expect:

• offload is the knob that decides whether this node runs on your card at all. disabled parks everything on the GPU (~39GB VRAM), model swaps components to CPU when idle (~16GB), aggressive adds single-frame VAE decode (~12GB). If you don't have a 4090 or better, this is the setting you'll live in.
• precision: fp16 is the sane default - fast and light. bf16 if you're on a newer card that likes it. fp32 doubles VRAM for a marginal quality bump you will not see on video.
• attention: sdpa (PyTorch native) is the default and always works. xformers and the sageattn variants only appear in the dropdown if you've installed those libraries yourself - the pack won't force them on you. If the list looks short, that's expected, not a bug.

Internally it's doing more than a weight load: it pre-computes the negative prompt embedding so sampling doesn't repeat that work, sets up a cosine-interp noise schedule tuned for STAR, and registers a star model folder so you can drop .pt weights in by hand if you'd rather not re-download. The output is one star_model connection, and it plugs straight into STAR Video Super-Resolution's star_model input. That's the only thing it connects to.

Install

ComfyUI Manager: search ComfyUI-STAR and install. Or by hand:

cd ComfyUI/custom_nodes
git clone --recursive https://github.com/ethanfel/Comfyui-STAR.git
cd Comfyui-STAR
pip install -r requirements.txt

Forget the --recursive? The pack auto-runs git submodule update --init on first import, so it recovers on its own - it even applies a small patch to the STAR submodule to fix OpenCLIP batch-first handling. Requirements are modest (diffusers, open-clip-torch, torchsde, einops, huggingface_hub); torch itself comes from your ComfyUI environment, and xformers/SageAttention are only needed if you want those backends.

Gotchas

The checkpoint load prints "missing keys" warnings because it loads with strict=False - usually benign, since some I2VGen keys don't exist in STAR. The real gotcha is the download: first load needs a good connection and patience, and offload defaults to disabled, which OOMs anything under ~40GB. Set offload to model before you hit Run, not after.