STAR Model Loader

Every other STAR node in this pack is quiet until this one loads. STAR Model Loader is the gate: it pulls down the actual video super-resolution checkpoint - the multi-gigabyte one - and hands it to the rest of the graph as a single "STAR model" object. If your first run sits there for a while with nothing happening, this is the node doing it.

STAR is a diffusion model for upscaling video (from NJU-PCALab's STAR project), and it ships two personalities you pick here: Light Degradation and Heavy Degradation. They're two separately-trained checkpoints for different source quality. If your video is merely small or soft - 720p you want cleaner and bigger - Light is your model. If it's genuinely battered - VHS transfers, 90s cable TV, brutal streaming compression - Heavy is the one trained to meet that level of damage, and it's the variant the community's early experiments liked most on old footage. Pick the one that matches your source, not your ambition. Point the Heavy model at a clean clip and it can over-correct into something that looks processed.

How it works

On first use for a given model type, the node auto-downloads the checkpoint from Hugging Face - SherryX/STAR hosts light_deg.pt and heavy_deg.pt - into ComfyUI/models/STAR/. They're roughly 2–3GB each, so budget disk and internet for the first run. Loading a checkpoint also pulls in two companions the model needs that the README lists as third-party pieces: the Stable Video Diffusion temporal VAE and the OpenCLIP text encoder, both fetched from Hugging Face automatically on that first model load. Nothing here is manual; everything here is slow once.

Two settings only:

• model - Light or Heavy Degradation, as above.
• precision - fp16 (default) or fp8 for the UNet weights. The tooltip is honest: FP8 needs compatible PyTorch and hardware. If you're not sure, leave fp16 - the fp8 path is a VRAM saver for people who know their stack supports it, not a free lunch.

The output is STAR model (type STARVSR_MODEL), which you wire to STAR Text Encode, STAR Sample, and STAR VAE Decode. That's the whole point of splitting it out: the model loads once, and ComfyUI caches it so you can tweak the prompt, the seed, the CFG, everything downstream, and never reload these gigabytes again. It also re-checks the file on disk, so swapping a checkpoint in models/STAR/ invalidates the cache correctly rather than silently running a stale model.

Installing the pack

Search ComfyUI-STARWrapper in ComfyUI Manager and install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/vjumpkung/ComfyUI-STARWrapper
cd ComfyUI-STARWrapper
pip install -r requirements.txt

Then restart ComfyUI. The honest caveats before you get excited: this pack is written against the newer V3 node API (comfy_api.latest), so you need a current ComfyUI build, not a year-old one. It also drags in a real dependency list - diffusers, open-clip-torch (pinned to 2.20.0), av, opencv-python, torchsde, fairscale, transformers, and a numpy<2.3 pin that can fight an environment stuck on numpy 2.x. And it wants a CUDA PyTorch plus an xformers that matches it; the pack ships an install script that tries to pick a matching xformers build, because the README warns CUDA 13 xformers does not work. This is not a three-second install, but it's a one-time one.

Troubleshooting

• First run downloads forever - that's the 2–3GB checkpoint plus VAE plus text encoder all landing at once. Let it finish once; it won't do it again.
• Run hangs on load or OOMs later - STAR is a genuinely heavy diffusion model; this node can't fix your VRAM. The rest of the pack chunks aggressively, but a 3090/4090-class card is the realistic floor for real work.
• Switching Light ↔ Heavy reloads - the model cache holds one type; flipping the dropdown is another slow load, so don't bounce between them while iterating.

If the load succeeds, your next stop is STAR Text Encode. The hard part of this pack isn't wiring - it's the wait.