Results do not match between demo and pretrained models

[ThangPM]

Hi,

I was trying to reproduce results I got from the demo (https://owlvit-tgc6as3cga-ew.a.run.app/) by loading the pretrained models: owlvit-base-patch32, owlvit-base-patch16 and owlvit-large-patch14 and following the usage guide on HuggingFace (https://huggingface.co/google/owlvit-base-patch32) for text-conditioned detection, but I could not get the same results. Do you have any suggestions for this discrepancy?

Best, Thang

[MaxTeselkin]

I noticed the same issue - especially in reference image mode. Huggingface integration is broken. Use scripts from original repo instead