Pangram Launches Image 1.0 and Flash 1.0 for AI Media Detection

Pangram has released Pangram Image 1.0 and Pangram Image Flash 1.0, moving its AI image and video detection models out of research preview and into a full product release. The launch matters because synthetic images and videos are becoming harder to distinguish by eye, and Pangram is positioning these models for platforms, publishers, marketplaces and other services that need an automated check before content is trusted or published.

According to Pangram, the release follows a research preview launched a few months earlier. In its announcement, written by founding computer vision engineer Isaac Kasahara and dated Nov 4, 2026, Pangram says community feedback helped expose weaknesses in the preview and set a higher bar for the new models. The company now presents Image 1.0 as its flagship image detector and Flash 1.0 as a smaller, faster model focused on video.

What is new in Pangram Image 1.0 and Flash 1.0?

Pangram Image 1.0 is the larger image-focused model. Pangram says it is designed for frontier AI image generators and adds heatmap output, which is intended to show where AI content appears inside an image instead of returning only a single yes-or-no judgement for the whole file. That distinction is practical: a photograph may be mostly real while one object, background area or edited region was generated or altered, and a heatmap gives a reviewer a place to inspect more closely.

Pangram Image Flash 1.0 is the smaller and faster model, and it is aimed primarily at video. According to Pangram, Flash 1.0 analyses video frame by frame, supports native video upload in the product, and is also available through an updated API that covers both image and video detection. Pangram says the image model has also been integrated into its Chrome extension, where it can flag AI-generated material while a user scrolls through social feeds, shopping pages, rental listings and other sites.

The company also says Image 1.0 is trained to remain reliable after common transformations. That includes compression, downscaling and other edits, as well as blur and noise that could be added to make synthetic content look more like a camera capture. Pangram presents robustness as a central design goal rather than an afterthought, because images shared online are rarely left in their original form.

How well does Pangram Image 1.0 perform on images in Pangram's tests?

Pangram describes its main image evaluation as a set of synthetic mirrors built from 100 real photographs. For each real photo, the company says it used 12 frontier image generators with prompts closely matching the original, producing 1,200 synthetic images in total. On that evaluation, Pangram reports 99.96% accuracy for Image 1.0 and says it compared the model with several commercial and academic detectors.

The mirror method is important to how Pangram frames the test. By starting from a real image and asking a generator to recreate something closely matched to it, the comparison is meant to be harder than simply asking whether a detector can spot an obviously synthetic picture. Pangram says this approach also reduces the risk that a detector simply learns to label certain scenes, subjects or styles as AI, because the real and synthetic versions share much of the same content.

Beyond its own mirror test, Pangram says it compared Image 1.0 with academic benchmarks including X-AIGD, Mirage-Test, AIGenImages2026, a Synthbuster and RAISE comparison, and Community Forensics. The blog post does not publish a full table of scores for each benchmark in the text supplied, so readers should treat the headline 99.96% figure as the result of Pangram's specific 1,200-image evaluation rather than a universal guarantee for every generator, camera type or editing pipeline.

How does Flash 1.0 handle AI video, and what results does Pangram report?

Flash 1.0 is Pangram's answer to a problem the company highlights with reference to a recent academic study called DF26. According to Pangram, that study found people were only slightly better than chance at distinguishing real videos of talking people from AI-generated ones, with existing detectors not performing much better. Pangram says Flash 1.0 scores 98.7% on DF26, which it describes as a 29-point improvement over the next best detector in that comparison.

Pangram also describes an industry comparison built from 129 real social media videos. It says 14 frontier video generators each produced 100 AI mirrors sampled from those real videos, seeded with the first real frame and prompted using a description of the video generated by a vision-language model. That produced 1,529 videos in total, which Pangram compared across multiple major AI video detection services. Again, the mirror design is meant to make the task difficult: the synthetic version starts from the same opening frame and scene, so a detector cannot rely only on broad differences in subject or setting.

For academic video benchmarks, Pangram names AIGVDBench, DF26 and Ivy-Fake, and says Flash 1.0 achieves state-of-the-art results across multiple benchmarks that together cover almost every frontier AI video generation model. As with images, the strongest claims are tied to the benchmarks and mirror sets Pangram selected. Independent replication, performance on newer generators and behaviour on heavily edited or unusual footage will be the real test as the models are used in the wild.

What are the practical limits for smaller images and edited media?

One concrete limit Pangram discloses concerns image size. The earlier release required a minimum input of 512 by 512 pixels because accuracy dropped at lower resolutions. With Image 1.0, Pangram says the minimum falls to 256 by 256 pixels. However, for images below 512 pixels, Pangram says detection results carry a reduced-confidence disclaimer. That means thumbnails, heavily downscaled screenshots and small profile or product images should not be treated with the same confidence as full-size uploads.

Robustness also cuts both ways. Pangram says it trains for edits, compression and low resolutions, and reports that the model remains robust to those transformations. Even so, detection is a probabilistic judgement, not proof of origin. A high score may justify closer review, labelling, a request for provenance information or a hold on publication, while a low score should not be read as absolute proof that media is authentic. Pangram's own framing around heatmaps supports that more careful workflow: localisation helps a human decide where to look next.

The Chrome extension and API also change how the models might be used. In a browser, the emphasis is on quick, contextual warnings during browsing. Through the API, the emphasis shifts to integration: a platform could check uploads at scale, a newsroom could route suspicious items for review, or a marketplace could add a signal to its moderation queue. In both cases, Pangram is selling detection as an input to a decision, not as a replacement for editorial judgement, platform policy or forensic investigation.

How could creators use detection alongside Grafy?

Grafy does not offer a Pangram-style AI detection model, and this launch should not be read as a Grafy feature. The genuine connection is workflow: creators who generate images or video in Grafy Creator may need to check provenance, label synthetic work correctly, or review whether an AI-edited asset could be mistaken for a real photograph or clip. A detector can sit beside the creation process as a separate verification step, especially when a project involves documentary material, product photos, rental or marketplace listings, or any context where audiences expect real media.

The same consideration applies in Grafy Film Studio, where AI-assisted documentary and short-film work can mix generated scenes with real footage. Before publishing, a creator may want to keep clear records of which shots were generated, which were captured, and how each was edited. Detection tools can help flag inconsistencies, but production records remain the stronger source of truth for a creator's own work.

For research and verification, readers can use Grafy Reader to read and question papers, PDFs and benchmark reports, including the academic detection studies Pangram cites. That pairing is useful: detection scores are easier to interpret when you can inspect the underlying evaluation design, datasets and limitations rather than relying on a headline percentage alone.

Frequently asked questions

Is Pangram Image Flash 1.0 the model Pangram positions for AI video detection?

Yes. According to Pangram, Flash 1.0 is the smaller, faster model built particularly for frontier AI video detection, with frame-by-frame analysis. Pangram says it also supports native video upload and API access for video, while the larger Image 1.0 model is positioned for image detection.

Did Pangram Image 1.0 lower the minimum accepted image size compared with the earlier release?

Yes. Pangram says the earlier research preview limited inputs to at least 512 by 512 pixels because accuracy fell at smaller sizes. With Image 1.0, Pangram says that minimum drops to 256 by 256 pixels. For images below 512 pixels, Pangram says results include a reduced-confidence disclaimer, so smaller inputs should be interpreted more cautiously.

Does Pangram Image 1.0 return a heatmap showing where AI content appears in an image?

Yes. According to Pangram, Image 1.0 provides heatmap output intended to localize AI content within a picture, rather than only returning a single overall verdict. Pangram illustrates this with a real photograph containing an AI-generated element, saying the heatmap can point to the affected area. That localization may help reviewers inspect specific regions before deciding how to act.

Source: Pangram, Introducing Pangram Image 1.0

---

Source: Hacker News — front page — Introducing Pangram Image 1.0