Pangram Image 1.0 claims state-of-the-art AI image and video detection with heatmaps

Pangram ships Image 1.0 and Flash 1.0

Pangram has taken its image detector out of research preview and released version 1.0, alongside a faster sibling model aimed at video. In a company blog post that reached the front page of Hacker News, Pangram presents Image 1.0 as its flagship detector for AI-generated images, while Pangram Image Flash 1.0 targets AI-generated video with frame-by-frame analysis. Both are available through the company's web tool, a Chrome extension, and an updated API.

Pangram says feedback from the earlier preview, including users probing its weaknesses, pushed the quality of the final release.

What the image model claims

To measure detection performance, Pangram built a test set by taking 100 real photographs and regenerating each one with 12 frontier image generators, using prompts written to closely match the originals — 1,200 synthetic images in total. According to Pangram, Image 1.0 scored 99.96% accuracy on that set, ahead of the commercial and academic detectors it compared against.

The model also outputs a heatmap showing where generated content sits inside an image, which matters for hybrid cases such as a genuine photograph containing an AI-inserted element. Pangram says it trained specifically against the transformations that often defeat detectors — recompression, downscaling, blur and added noise — and reports strong results across academic benchmarks including X-AIGD, Mirage-Test, AIGenImages2026, a Synthbuster-plus-RAISE comparison, and Community Forensics.

There is one practical trade-off: the minimum input size drops from 512x512 pixels to 256x256, but results below 512 pixels now come with a reduced-confidence disclaimer.

Video detection and the DF26 benchmark

For video, Pangram points to a recent academic study called DF26, which found that humans performed close to chance when separating real footage from AI-generated videos of people talking, and that modern automated detectors did not fare much better. On DF26, Pangram says Flash 1.0 reaches 98.7%, a 29-point lead over the next-best detector.

For an industry comparison, the company took 129 real social media videos and produced AI counterparts using 14 frontier video generators, seeding each generator with the real video's opening frame and a machine-generated description of the action, for 1,529 videos in total. Because each real video and its synthetic counterpart share the same scene and subject, Pangram argues the setup tests genuine discrimination rather than a bias toward labelling certain kinds of content as fake.

Video upload is now natively supported: Flash 1.0 analyzes uploaded footage frame by frame and reports which parts of it come back as AI-detected.

Chrome extension and API

The Chrome extension now runs Image 1.0 in the browser, flagging synthetic media as you scroll social feeds, shopping sites, rental listings, and other places where knowing content is authentic matters, according to Pangram. The company's API has also been updated to cover both image and video detection.

Why it matters

Every headline number here is self-reported, drawn from test sets Pangram assembled itself, and the adversarial back-and-forth between generators and detectors means any current lead can erode as new models ship. Even so, the release addresses a real and widening gap: per the DF26 findings Pangram cites, neither people nor existing tools reliably tell real footage of talking humans apart from generated video. Detection with localization — showing where in a file the synthetic material appears, rather than issuing a single binary verdict — is especially useful for edited and hybrid media that pure classifiers miss. And by shipping a browser extension and an API alongside the web tool, Pangram is pushing the capability out to where the content actually circulates, not just to researchers running benchmarks.