ZoomCap routes image evidence into richer captions
In plain words
ZoomCap is a method for writing detailed image captions that can be used to train vision-and-language models. It makes a draft, looks for missing details, and gathers more evidence from relevant image regions or other sources before adding details that fit the draft and are not repeats.
The authors’ FactCycle test found more claims per image that were both supported by the original image and present in a reconstruction made from the caption than with ScaleCap captions. That result applies to the test’s selected image types, not every kind of image.
The research
Image captions are often used to teach vision-and-language models, but length and fluency do not show whether visual details survived or whether claims are supported.
The paper asks whether a captioning method can gather more specific evidence while keeping the resulting claims verifiable and recoverable from the caption.
ZoomCap routes new observations from a draft caption toward image gaps, then filters details so captions can carry more supported, recoverable evidence into training.
ZoomCap begins with a draft caption and directs further evidence gathering toward details the draft may have missed. The authors test whether this produces claims that are both grounded in the image and recoverable from the caption alone, rather than merely longer descriptions.
In FactCycle, independent judge pairs checked caption claims against the source image and a reconstruction generated from the caption. ZoomCap-450K produced an average of 52.4 claims per image supported at both anchors, versus 31.5 for ScaleCap-450K on the shared 250-image pool. The result supports more recoverable detail in that evaluation.
FactCycle tested 250 images from five scene categories and excluded screenshots, documents, charts, collages, artwork and illustrations, so it did not test ZoomCap’s OCR branch. The medical results are benchmark evaluations, not clinical validation.