Pangram has released Pangram Image 1.0, the successor to the research preview it launched a few months ago. The company says the model delivers state-of-the-art performance in both AI image and AI video detection. The flagship model features heatmap output that shows exactly where AI content appears in an image, along with robustness to edits, compression, and low-resolution inputs.
Pangram also introduced Pangram Image Flash 1.0, a smaller and faster model that focuses on frontier AI video generators. It analyzes AI-edited videos frame by frame and is available through a Chrome extension for checking content in a social feed, as well as through API support for both images and videos.
In internal testing, Pangram Image 1.0 was evaluated on 1,200 synthetic images created with 12 frontier AI image generators, each closely matching one of 100 real photos. The model reached 99.96% accuracy on this set. The company also reports that the model was tested against several academic benchmarks, including X-AIGD, Mirage-Test, AIGen Images2026, Synthbuster with RAISE, and Community Forensics.
The update also lowers the minimum input resolution from 512x512 to 256x256, since the new model holds accuracy better at smaller sizes. Results below 512 pixels will carry a reduced-confidence disclaimer. For video, Pangram cited a recent academic study, DF26, which found humans barely outperform random guessing when distinguishing real talking-person videos from AI versions. Flash scores 98.7% on DF26, a 29-point gain over the next best detector.
Source: Hacker News · Summarized by HeadlinesBriefing