textiled

SAM-based parquet plank grid pipeline: normalize lighting → Meta SAM 3 masks → fit a 2D lattice and write debug images / JSON.

Setup

uv sync

Caveats: NOTES.md.

SAM3 weights (facebook/sam3)

Ultralytics loads sam3.pt from Meta’s Hugging Face repo facebook/sam3. This pipeline resolves weights in order: a file named sam3.pt next to your working directory (see SAM_WEIGHTS in parquet_sam_grid.py), then the Hugging Face hub cache. It does not download from the Hub automatically—you must fetch the file once.

1. Access — Open facebook/sam3, sign in, and accept any license / request access if the card says the model is gated.
2. Token — Create a read token at huggingface.co/settings/tokens.
3. Authenticate (pick one):

• uv run hf auth login and paste the token, or
• export HF_TOKEN=<your_token> in the shell where you download and run.

4. Download (pick one; run from the repo root so --local-dir . lands beside SAM_WEIGHTS):

• Into the project folder (matches the default sam3.pt filename):
• Or only into the Hub cache (no sam3.pt in the repo; resolution still finds it under ~/.cache/huggingface/hub unless you override cache dirs):

  hf download facebook/sam3 sam3.pt

Optional: HF_HOME / HF_HUB_CACHE change where cached snapshots live; see hf_hub_cache_dirs() in sam/backend.py.

Run

python3 src/parquet-sam-grid.py --image samples/herringbone.png --text "wooden parquet rectangular planks"

Outputs go to outputs/<image_stem>/. Weights: SAM_WEIGHTS in parquet_sam_grid.py (default sam3.pt + hub resolution).