CLIP's accuracy is significantly lower on images from blind/low vision users compared to web-crawled images due to sensitivities to image content, quality, and text content.
