99.2% of synthetic images detected: the results announced by Pangram Image 1.0

Pangram deploys two AI-generated image and video detection models, with a Chrome extension and visual analysis of suspicious content.

An authentic photograph can contain an element entirely generated by artificial intelligence. A synthetic video can take the foreground of a real recording, then reproduce the same movements and the same environment. These hybrid contents make it harder to distinguish between captured images and fabricated images. To identify them, Pangram is launching two new detection models capable of analyzing photographs, locating suspicious regions, and examining videos frame by frame.

Already specialized in recognizing text generated by AI, the American company is expanding its activity with Pangram Image 1.0 and its lighter version, Image Flash 1.0. According to the results published on November 4, 2026, the first detects 99.2% of synthetic images under the evaluation conditions used, with a false positive rate of 0.27%. In other words, about 27 authentic photographs out of 10,000 would be incorrectly classified as AI-generated in this test framework.

These results are reported by Pangram and have not yet been the subject of documented independent replication. Spotting changes in a real photograph

One of the most concrete new features concerns the localization of generated elements. Rather than providing only a global classification, the detector can display a heat map indicating the regions with characteristics associated with artificial generation.

A restaurant photograph containing a mural created by AI is notably used as an example in the official demonstration. The system highlights the suspicious area while distinguishing the rest of the photographed scene.

This function addresses a specific problem: an image is not necessarily entirely real or synthetic. Generative editing tools can modify an object, replace a background, or add elements while preserving most of the original photograph.

The ability to localize interventions could interest teams responsible for verifying visual content, but the heat map remains a model estimate. It does not constitute irrefutable proof that a specific region was generated.

The detector was also trained to better withstand common transformations on the internet. Compression, reduced resolution, blur, and added noise can alter the characteristics used to recognize a synthetic image. These transformations occur both when publishing on social networks and in deliberate attempts to bypass a detection system.

Since its first experimental version released in July, the minimum accepted image size has increased from 512 × 512 to 256 × 256. Below 512 × 512, the results are accompanied by a warning indicating reduced reliability. Synthetic images built from real photographs

To evaluate its model, Pangram assembled a test set from 100 authentic photographs. Twelve recent image generators then produced 1,200 synthetic versions, using prompts designed to faithfully reproduce the original scenes.

On this set, the company reports an accuracy of 99.96%. This figure differs from the 99.2% detection rate announced for the product: the two measures correspond to different evaluations and different indicators, which should not be confused.

The protocol is particularly noteworthy because it avoids comparing only real and artificial images depicting completely different subjects. By reproducing the same scenes, the researchers aim to reduce the risk that the detector relies on content differences rather than generation features.

The results are also compared with several academic evaluation sets, including X-AIGD, Mirage-Test, AIGenImages2026, and Synthbuster. The performance presented places the model ahead of several competing solutions, but these comparisons remain conducted and published by Pangram.

The accuracy obtained on a photograph set built for evaluation does not guarantee identical performance on all images distributed online. New generative models, complex edits, and photographic processing can change the detection conditions. A lighter model for examining videos

Video analysis is the second part of this update. With Image Flash 1.0, Pangram introduces a faster model designed to examine the frames of a sequence one after another and identify segments that may have been generated or modified.

For its commercial comparisons, the company selected 129 authentic videos from social networks. Fourteen generators then produced 1,400 artificial videos, using the real foreground and a description of the scene as references.

The evaluation therefore covers 1,529 videos. Pangram announces an accuracy of 98.57% for Image Flash 1.0, compared with 84.2% for the best-ranked competitor in its comparison. The method aims to confront detectors with sequences sharing similar subjects and compositions.

The result should nevertheless be interpreted with caution. Artificial videos are far more numerous than authentic videos in this set. Without a detailed examination of errors by category, a global score is not enough to determine how often real videos would be incorrectly flagged.

Another result concerns DF26, an academic benchmark dedicated to synthetic videos. According to Pangram, the model achieves a score of 98.7%, or about 29 points above the best cited reference system.

This comparison requires additional caution: scientific evaluations of DF26 use AUROC in particular, an indicator that measures the ability to distinguish two categories across different thresholds. It does not directly correspond to a percentage of videos correctly classified. Pangram’s communication does not specify the calculation methods sufficiently to equate this score with 98.7% accuracy in everyday use. Detection integrated directly into browsing

The company is not reserving these capabilities for a specialized analysis tool. The image detection model is now integrated into its Chrome extension, intended to flag potentially synthetic content while browsing.

The uses mentioned cover social networks, commercial listings, sales platforms, and real estate rental sites. Users can thus check visuals without necessarily saving them