Checkpoints¶

Strands Sapiens looks for Sapiens2 weights on disk. It doesn't download them for you (model licensing varies - see MODEL_ZOO.md upstream).

Checkpoint root¶

The root directory is controlled by the SAPIENS_CHECKPOINT_ROOT env var. Default: ~/sapiens2_host.

export SAPIENS_CHECKPOINT_ROOT=/data/sapiens2_host

Expected layout¶

$SAPIENS_CHECKPOINT_ROOT/
├── pretrain/
│   ├── sapiens2_0.1b_pretrain.safetensors
│   ├── sapiens2_0.4b_pretrain.safetensors
│   ├── sapiens2_0.8b_pretrain.safetensors
│   ├── sapiens2_1b_pretrain.safetensors
│   ├── sapiens2_1b_4k_pretrain.safetensors
│   └── sapiens2_5b_pretrain.safetensors
├── seg/
│   ├── sapiens2_0.4b_seg.safetensors
│   ├── sapiens2_0.8b_seg.safetensors
│   ├── sapiens2_1b_seg.safetensors
│   └── sapiens2_5b_seg.safetensors
├── normal/    (same sizes as seg)
├── albedo/    (same sizes as seg)
├── pointmap/  (same sizes as seg)
├── pose/      (same sizes as seg)
└── detector/
    └── detr-resnet-101-dc5/              # required for pose (HuggingFace model dir)

Supported sizes per task¶

Task	Sizes
`pretrain`	`0.1b` · `0.4b` · `0.8b` · `1b` · `1b_4k` · `5b`
`seg`	`0.4b` · `0.8b` · `1b` · `5b`
`normal`	`0.4b` · `0.8b` · `1b` · `5b`
`albedo`	`0.4b` · `0.8b` · `1b` · `5b`
`pointmap`	`0.4b` · `0.8b` · `1b` · `5b`
`pose`	`0.4b` · `0.8b` · `1b` · `5b`

The 1b_4k variant uses the 1B-parameter backbone at a 4096×3072 input resolution.

Discover what you have¶

from strands_sapiens import sapiens_info
print(sapiens_info())

{
  "status": "success",
  "checkpoint_root": "/data/sapiens2_host",
  "checkpoint_root_exists": true,
  "available": {
    "pretrain": ["0.1b"],
    "seg":      ["0.4b"]
  },
  "detector_present": true,
  "cuda": {"available": true, "device_count": 1, "device_name": "Orin"},
  "sapiens_package": true
}

Download sources¶

Model weights: see Sapiens2 MODEL_ZOO.md - typically hosted on Hugging Face under facebook/sapiens2-*.
Person detector (DETR): huggingface-cli download facebook/detr-resnet-101-dc5 --local-dir $SAPIENS_CHECKPOINT_ROOT/detector/detr-resnet-101-dc5

Size vs. VRAM cheat-sheet¶

Size	Params	Min VRAM (fp16 inference)	Notes
0.1b	100M	~1 GB	Pretrain only
0.4b	400M	~2 GB	Good default for Jetson
0.8b	800M	~4 GB
1b	1B	~6 GB
1b_4k	1B	~16 GB	4096×3072 resolution
5b	5B	~24 GB	Best quality

(Values approximate - actual usage depends on batch size & precision.)