Pangram introduced Image 1.0 and smaller Image Flash 1.0 on November 4, adding video analysis to its image detector.
The big change
- Image 1.0 produces a heatmap of suspected AI-generated regions.
- Flash checks video frame by frame.
- Results are signals for investigation, not proof of origin.
From images to video
The company says its image model accepts inputs down to 256 by 256 pixels; results below 512 carry a reduced-confidence notice. Its Chrome extension checks images; its API supports image and video. Video uploads receive frame-level analysis.
What the tests say
Pangram reports 99.96% accuracy on an image set using 100 photographs and 1,200 synthetic counterparts from 12 generators. Its video comparison used 129 real clips and 1,400 generated clips from 14 generators; it reports a 98.7% score on DF26. The post does not detail the image accuracy calculation in its text. These are company-run evaluations.
A false positive is a real image incorrectly labeled AI-generated. In its July preview, Pangram reported a 0.16% false-positive rate on 10,000 pre-2022 ReLAION images. The result concerns the earlier preview.
Independent Community Forensics research identifies generalization to previously unseen generators as a central detection challenge. Its authors studied detector training across a large collection of generators. In consequential settings, checking original files and provenance alongside any detector result is more defensible than treating one score as a verdict.
