ComfyUI Node
STAR Model Loader
Loads the STAR video super-resolution model (UNet+ControlNet, OpenCLIP text encoder, temporal VAE). All components are auto-downloaded on first use.
STAR Model Loader
- star_model
◄model_name▾►
◄precisionfp16►
◄offloaddisabled►
◄attentionsdpa►
CategorySTAR
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| model_name | COMBO | STAR checkpoint to load. light_deg for mildly degraded video, heavy_deg for heavily degraded video. Auto-downloaded from HuggingFace on first use. | |
| precision | COMBO | fp16 | Inference precision. fp16 is recommended (fastest, lowest VRAM). bf16 for newer GPUs. fp32 disables autocast for maximum quality at 2x VRAM cost. |
| offload | COMBO | disabled | disabled: all on GPU (~39GB). model: swap UNet/VAE/CLIP to CPU when idle (~16GB). aggressive: model offload + single-frame VAE decode (~12GB). |
| attention | COMBO | sdpa | Attention backend. sdpa: PyTorch native (default, always available). xformers: original backend. sageattn: SageAttention auto-select. Other sageattn_* variants: specific SageAttention kernels for fine-tuning speed/precision. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| star_model | STAR_MODEL | — |