Pangram launches Pangram Image 1.0 AI detector with heatmaps, plus Flash for video

Pangram has announced Pangram Image 1.0, which it calls its new flagship AI image detection model. It follows a research preview the company released "a few months ago". Alongside it comes Pangram Image Flash 1.0, a smaller, faster model aimed at detecting AI-generated video. Pangram thanks the community for trying the preview and finding its weaknesses.

The company lists what is new. Pangram Image 1.0 is billed as state-of-the-art at detecting frontier AI image generators, outputs a heatmap that localizes where AI appears in an image, and is described as robust to edits, compression and low resolutions. Flash 1.0 is billed as state-of-the-art at detecting frontier AI video generators and analyzes AI-edited video frame by frame. Both models are also available through the Pangram Chrome extension and the API, which now supports image and video.

On images, Pangram says it tested the model against several commercial and academic detectors. For its own evaluation set it made synthetic "mirrors" of 100 real photos using 12 frontier AI image generators, with prompts that closely matched the original photo, giving 1,200 synthetic images. Pangram Image 1.0 reaches 99.96% accuracy on this set. The company also compared the model on academic image benchmarks, naming X-AIGD, Mirage-Test, AIGenImages2026, Synthbuster + RAISE and Community Forensics as sources. The comparison charts are not reproduced in the text.

On robustness, Pangram notes that social platforms compress and downscale images, and that bad actors can add blur, noise and other artifacts to fool detectors. It says it specifically trains on these transformations and has found the model robust to them. On resolution, the initial release required images of at least 512x512 because accuracy dropped at smaller sizes. The limit is now 256x256, since the new model is more accurate at low resolutions. Below 512, results come with a reduced confidence disclaimer.

For video, Pangram cites a recent academic study, DF26, which showed that humans are barely better than random at telling real videos of people talking from AI ones, and that modern detectors were not much better. On DF26, Flash 1.0 scores 98.7%, which Pangram calls a 29-point leap over the next best detector. For an industry comparison, it took 129 real social media videos and generated AI mirrors with 14 frontier video generators, 100 videos each, sampled from the 129. Each generator was seeded with the first real frame and prompted with a description of the video written by a vision-language model, to make a hard synthetic mirror. Multiple major AI video detection services were then compared across the 1,529 total videos. Pangram says the mirror design avoids bias toward calling certain content types AI, because each real video and its mirror share the same scene and subject. It also reports state-of-the-art results across academic video benchmarks (AIGVDBench, DF26, Ivy-Fake).

Pangram now accepts video uploads directly. Flash 1.0 analyzes the video frame by frame and shows which parts come back as AI detected. The Chrome extension flags AI-generated content while users scroll social media, online shopping, rental listings and other sites. The post is written by Isaac, a research engineer working on AI image and video detection at Pangram.

Key facts

Pangram Image 1.0 is a new AI image detector with a heatmap output that localizes where AI content appears in an image.
Pangram reports 99.96% accuracy on its own evaluation set: 100 real photos plus 1,200 synthetic mirrors made with 12 image generators.
The minimum image input size drops from 512x512 to 256x256; results below 512 carry a reduced confidence disclaimer.
Pangram Image Flash 1.0 handles video frame by frame and scores 98.7% on the DF26 benchmark, which Pangram calls a 29-point lead over the next best detector.
Both models are in the Pangram Chrome extension and the API; the benchmark results are the company's own.

Why it matters

Pangram argues that telling real from AI-made media is getting hard for people and tools alike. It cites the DF26 study, which found humans barely better than random at spotting AI videos of people talking, and detectors not much better. Heatmap localization is the practical step here: instead of a single verdict, the model marks where in a photo the AI content sits, such as an AI element inside an otherwise real photograph. Extending the same product to video and lowering the image size limit widens the range of media it can check.

Who it affects

The people Pangram points to are anyone browsing social media, online shopping or rental listings who wants to know whether what they see is real, through the Chrome extension. Developers can reach the models through the API, which now handles image and video. Platforms and bad actors come up too: Pangram notes that platforms compress and downscale images and that malicious users add blur or noise to try to fool detectors.

How to use it

There is a web tool (the post says "Try it here"), the Pangram Chrome extension, which flags AI-generated content as you browse, and the API, which supports image and video. Video can be uploaded directly, and Flash 1.0 returns a frame by frame analysis showing which parts are detected as AI. Images should now be at least 256x256; below 512x512 expect a reduced confidence disclaimer. No pricing, rate limits or API terms are given in the post.

How solid is it

All the figures are Pangram's own, from a first-party announcement. The one image accuracy number in the text is 99.96% on its 1,200-image synthetic set paired with 100 real photos; no false-positive rate or breakdown by generator is given. The per-detector scores for competitors are in charts that are not in the text, and no competitor is named. No numeric result is given in the text for the academic image benchmarks, the video industry comparison or the robustness tests. The generators used are not named, and no independent or third-party verification is mentioned. The DF26 result (98.7%, a 29-point margin) is stated by Pangram.

Risks and caveats

Treat the headline numbers as vendor claims until others reproduce them. The 99.96% figure comes from a small set built by the company itself, with prompts matched to the original photos. At resolutions below 512, Pangram itself attaches a reduced confidence disclaimer, and it says accuracy dropped off at lower resolutions in the first release. Robustness to compression and adversarial edits is asserted from the company's own training and testing. Because the text names no competitors and gives no scores for them, the claims of state-of-the-art performance cannot be checked from the post alone.

“We specifically train to detect these kinds of transformations, and we have found our model to be robust to these edits.”

— Pangram, announcement of Pangram Image 1.0