ComfyUI Node
BD ATR Human Parser
Parse a human image into 18 semantic regions using an ATR-scheme SegFormer. Default weights (mattmdjaga/segformer_b2_clothes) are MIT-licensed. Outputs a HUMAN_PARSE_MAP for the mask-split node.
BD ATR Human Parser
- image
- parse_map
◄model_idmattmdjaga/segformer_b2_clothes►
◄deviceauto►
◄dtypefp16►
◄confidence_threshold0.00►
◄cache_dir►
Category🧠BrainDead/Segmentation
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| model_id | STRING | mattmdjaga/segformer_b2_clothes | HuggingFace repo id (must use ATR 18-class scheme) |
| deviceopt | COMBO | auto | 3 options: auto, cuda, cpu |
| dtypeopt | COMBO | fp16 | 3 options: fp16, bf16, fp32 |
| confidence_thresholdopt | FLOAT | 0.000–1 | Pixels with max-softmax below this become background. 0 = off. Try 0.5–0.7 to clean noise. Tip: for sharper boundaries, downsize the image to ~512px before this node (ATR was trained at 512×512). Large inputs blur edges on the upsample. |
| cache_diropt | STRING | Override HF cache dir. Blank → HF_HOME env var, or ComfyUI models/huggingface/, or HF default. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| parse_map | HUMAN_PARSE_MAP | — |