The image release returns attack type, whether an edit is localized, and heatmap type in the API and Detect report. Manual edits such as splicing, text edits and redaction now count as fake, and document and ID forgery recall now exceeds 86%, up from under 30% before.