WOW-Seg-J

Jittor inference weights for WOW-Seg, converted from the official BF16 model.safetensors checkpoint.

Files

• wow_seg.pkl: Jittor checkpoint in FP16.
• wow_seg.manifest.json: tensor count and source SHA-256 manifest.
• config.json, tokenizer files, and preprocessor configuration: required by jittor_wowseg inference code.

The checkpoint contains 637 tensors and 938,193,024 parameters. The original PyTorch/Hugging Face checkpoint is not included.

Usage

Use the jittor_wowseg support code from the WOW-Seg-Meta repository:

python -m jittor_wowseg.infer \
  --model-dir path/to/WOW-Seg-J \
  --image path/to/image.jpg \
  --mask path/to/mask.png

The Jittor implementation defaults to 12 dynamic image tiles to match the original inference configuration. Reduce this only when GPU memory is limited.

Conversion

The release was produced with jittor_wowseg/convert_weights.py. It converts BF16 tensors to FP16 because FP16 is the portable Jittor CUDA format, preserves all original parameter keys, and validates the checkpoint before inference.