ComfyUI Node
LocateAnything Model Loader
Loads NVIDIA LocateAnything-3B from a local snapshot or Hugging Face. Missing files can be downloaded automatically into models/LocateAnything. The official checkpoint requires trust_remote_code=True.
LocateAnything Model Loader
- model
◄model_sourcenvidia/LocateAnything-3B►
◄download_modeltrue►
◄deviceauto►
◄dtypeauto►
◄attentioneager►
◄use_batch_runtimefalse►
◄runtime_attentionla_flash►
◄vision_attentionauto►
◄schedulerpipeline►
◄group_size0►
◄strict_attnfalse►
CategoryLocateAnything
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| model_source | STRING | nvidia/LocateAnything-3B | Hugging Face repo ID or local snapshot directory. |
| download_model | BOOLEAN | true | Download missing model files into models/LocateAnything. |
| device | COMBO | auto | Execution device. auto uses the device selected by ComfyUI; cuda is recommended when available. |
| dtype | COMBO | auto | Model precision. auto selects bfloat16 or float16 on CUDA and float32 on CPU. |
| attention | COMBO | eager | Eager is the conservative compatibility path. SDPA can be faster but may be less stable on some setups. |
| use_batch_runtime | BOOLEAN | false | Use the official hybrid batch runtime from the model snapshot when available. |
| runtime_attention | COMBO | la_flash | Attention backend for the optional official batch runtime. |
| vision_attention | COMBO | auto | Vision attention backend for the optional official batch runtime. |
| scheduler | COMBO | pipeline | Hybrid batch scheduler used by the official batch runtime. |
| group_size | INT | 00–1024 | Batch runtime hybrid grouping. 0 keeps the upstream default. |
| strict_attn | BOOLEAN | false | Require the configured batch attention backend instead of allowing runtime fallback. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| model | LOCATEANYTHING_MODEL | Loaded LocateAnything runtime shared with grounding nodes. |