Pangram claims major gains in spotting AI images and videos

Pangram has released two new detection models that it says can identify AI-generated images and videos, including content that has been compressed, edited, or mixed with real footage. In an official Pangram blog post, Isaac Kasahara writes that the company’s new flagship, Pangram Image 1.0, is designed for image detection and can also analyze AI video content.

The model produces a heatmap intended to show where AI-generated elements appear within an image. This could be useful for detecting manipulated photographs where only part of the image has been created or altered with generative AI. Pangram says Image 1.0 also performs better on lower-resolution files. It now accepts images as small as 256 by 256 pixels, although results below 512 pixels come with a reduced-confidence notice.

Company reports high benchmark scores

Pangram says it tested Image 1.0 on 1,200 synthetic images. The test set paired 100 real photographs with AI-generated versions created by 12 image generators. The company reports 99.96 percent accuracy on that evaluation. It also compares the model with other detectors on academic datasets including X-AIGD, Mirage-Test, AIGenImages2026, Synthbuster, RAISE, and Community Forensics.

The second model, Pangram Image Flash 1.0, is a smaller and faster system focused on video. Pangram reports a score of 98.7 percent on the DF26 benchmark, which evaluates AI-generated talking-head videos. The company says this result is 29 percentage points above the next best detector in its comparison.

For another test, Pangram used 129 real social media videos and generated synthetic counterparts with 14 video models. The company says this approach reduces the chance that a detector simply associates certain scenes or subjects with AI content. The results are company-reported and the post does not provide independent validation of the tests.

Pangram has also added native video uploads, frame-by-frame analysis, API support, and Chrome extension support for scanning content on websites.