weldvision-ensemble

WeldVision Ensemble

Four-model YOLO segmentation ensemble for weld-surface defect inspection.

Models

• weights/best.pt
• weights/best_v0.pt
• weights/crack_specialist.pt
• weights/spatters_specialist.pt

Inference runs locally after downloading the weights from the Hugging Face Hub. No Hugging Face Inference Endpoint is required.

Install

pip install ultralytics huggingface_hub opencv-python-headless numpy

Use

from weldvision import WeldVision

model = WeldVision.from_pretrained(
    "bhavibhatt/weldvision-ensemble"
)

result = model.predict("weld.jpg")

print(result)

The first call downloads and caches the four weights. Subsequent calls reuse the local Hugging Face cache.

Important

This is a proof-of-concept weld surface inspection system. The score and PASS/REVIEW/FAIL rules are prototype decision logic and are not welding-code acceptance criteria or a substitute for qualified human inspection.