Pangram Image 1.0 and Image Flash 1.0: AI Image and Video Detection Leaves Preview, With Heatmaps and Video Uploads
Three months after its research preview, Pangram's image detector now spots AI edits inside real photos, accepts smaller images, checks video frame by frame and runs in its Chrome extension. Every accuracy figure is still Pangram's own.
TL;DR
Pangram launched Pangram Image 1.0 and the smaller Pangram Image Flash 1.0 on 4 November 2026, replacing the research preview of its AI image detector released on 29 July. Image 1.0 adds a heatmap that shows where AI content sits in a real photo, is trained to resist compression, blur and noise, and lowers the minimum input size from 512x512 to 256x256. Image Flash 1.0 handles video: Pangram reports 98.7% on the DF26 academic benchmark, 29 points ahead of the next detector, and frame-by-frame analysis of uploaded videos. Image 1.0 scored 99.96% on a 1,200-image test set Pangram built itself, and no independent evaluation has been published yet.
Pangram built its name on detecting AI-written text. In July it added a research preview for images. This week's release makes image detection a full product and adds video, the area where Pangram says both people and existing detectors do worst.
What changed since the July preview
| Research preview (29 Jul) | Image 1.0 and Image Flash 1.0 (4 Nov) | |
|---|---|---|
| Models | One image model | Image 1.0 for images, Image Flash 1.0 (smaller and faster) for video |
| Smallest image accepted | 512 x 512 | 256 x 256, with a reduced-confidence note below 512 |
| Where AI appears | Whole-image verdict | Heatmap that localises AI content inside a real photo |
| Video | Some AI video generators covered (Kling, Seedance, Veo, Wan) | Dedicated video model, native upload, analysed frame by frame |
| Chrome extension | Coming soon | Image 1.0 flags AI images while you browse |
| Image API | Invitation only | Image and video detection |
The accuracy claims, and who ran them
- Images: Pangram made AI "mirrors" of 100 real photos with 12 frontier image generators, prompted to match each original, and Image 1.0 scored 99.96% on those 1,200 images. Pangram designed this test itself.
- The Next Web reports Pangram's figures as catching 99.2% of AI images while wrongly flagging 0.27% of real ones.
- Academic image benchmarks: Pangram compares Image 1.0 on X-AIGD, Mirage-Test, AIGenImages2026, Synthbuster plus RAISE, and Community Forensics, with results shown as charts.
- Video: Image Flash 1.0 scores 98.7% on DF26, a recent study of talking-head videos in which, Pangram notes, humans did barely better than chance. Pangram says that's 29 points ahead of the next-best detector.
- Industry comparison: 1,529 videos, made of 129 real social media clips plus AI mirrors from 14 video generators, each seeded with the real clip's first frame. Pangram doesn't name the competing services in the text.
What isn't in the announcement
- Prices: the post gives no API price per image or per second of video.
- Independent testing: none of the new models' results have been published by an outside group yet.
- Which video generators were used, and how each one scored.
What this means if you work with AI images and video
- For moderation, newsrooms and marketplaces, the heatmap matters more than the headline accuracy. A real photo with one AI-edited object is the hard case, and a whole-image yes or no can't show it.
- Treat detector scores as a signal, not proof. All the figures here come from Pangram, and the test sets were built to its own design.
- If you publish AI-generated images legitimately, expect more of them to be flagged by browser extensions like this one, and label them clearly.
Sources
- Pangram: Introducing Pangram Image 1.0 (official announcement, Nov 4, 2026)
- Pangram: Introducing Pangram Image Detection (Research Preview) (Jul 29, 2026)
- The Next Web: Pangram's new models spot AI fakes in images and video, it says (Nov 4, 2026)
- Unite.AI: Pangram Launches Two Models for AI Image and Video Detection
AI Industry Reporter
Priya covers model releases, industry announcements, and the gap between what labs claim and what independent evaluators actually find. She reads the primary source - the paper, the system card, the benchmark org's own statement - before writing a word.
More on AI Image Generators
Pruna's P-H3-Image Makes 8-Megapixel Images in 8 Steps, but Its License Excludes the US, EU, UK and South Korea
Priya Nair · 4 min
Canva Craft: Its First In-House Image Model Is Trained on Templates, Not Photos, and Plans the Layout Before It Draws
Priya Nair · 4 min
OpenAI Stops Taking ChatGPT Ads From Rival Image and Audio Generators. What Adobe Said, and the Questions OpenAI Hasn't Answered
Priya Nair · 4 min