ComfyUI Node
SceneCut Validator Loader
DINOv3 ViT-L/16 (models/clip_vision) and optionally SSCD (models/scenecut) for cut validation. Both download on first use; DINOv3 from Meta's gated repo if you are logged in to Hugging Face, else from a byte-identical mirror.
SceneCut Validator Loader
- validator
◄dinov3▾►
◄sscdsscd_disc_large.torchscript.pt►
◄precisionfp32►
◄devicedefault►
CategorySceneCut/advanced
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| dinov3 | COMBO | 1 options: dinov3_vitl16_pretrain_lvd1689m.safetensors | |
| sscd | COMBO | sscd_disc_large.torchscript.pt | Copy-detection descriptor used to break close calls and judge flashes. 'none' validates with DINOv3 alone. |
| precision | COMBO | fp32 | fp32 matches the calibration. bf16 is faster with tiny differences in the scores. (fp16 is not offered: DINOv3 overflows in it.) |
| device | COMBO | default | default: ComfyUI's GPU, managed with the other models. cpu: slow but needs no GPU. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| validator | SCENECUT_VALIDATOR | — |