kernel33

I tried object segmentation and it’s really good

Are you running everything through a single end-to-end vision model, or do you dynamically dispatch to specialized OCR, detection, and segmentation backends?