komorebi01/glls-abound-1shot

GLLS ABounD 1-shot Localizer Artifacts

This repository contains the ABounD 1-shot model artifacts used by GLLS for MVTec AD and VisA frontend runs. In GLLS, mvtec and visa with 1-shot route to this ABounD artifact; 0-shot and non-MVTec/VisA frontend runs use the AdaptCLIP fallback.

Directory Layout

glls-abound-1shot/
+-- README.md
+-- model_config.json
+-- model/
|   +-- ViT-L-14-336px.pt
+-- mvtec/
|   +-- final_memory_bank_mvtec.pt
|   +-- final_soft_prompt_state_mvtec.pth
|   +-- final_vvclip_model_state_mvtec.pth
+-- visa/
    +-- final_memory_bank_visa.pt
    +-- final_soft_prompt_state_visa.pth
    +-- final_vvclip_model_state_visa.pth

model_config.json records the artifact family, dataset routing, file paths, and calibrated image/pixel thresholds used by the localizer.

GLLS Setup

Download or place this repository at the expected GLLS model location, then set:

export GLLS_ABOUND_MODEL_PATH="/path/to/glls-abound-1shot/model"
export GLLS_ABOUND_SAVE_PATH="/path/to/glls-abound-1shot"

GLLS uses GLLS_ABOUND_MODEL_PATH for the ABounD backbone directory and GLLS_ABOUND_SAVE_PATH for the dataset-specific mvtec/ and visa/ state files.

License

License status is not confirmed in this artifact card. Verify the ABounD code, checkpoint, dataset, and backbone redistribution terms before public release, redistribution, or commercial use.