Pangram Image 1.0 review (2026): an honest verdict on the AI image and video detector — accuracy, heatmaps, video, robustness, limits, and who it's for.
Pangram Image 1.0 is one of the more credible AI-visual detectors released in 2026: a reported 99.96% on its own image test set, a video sibling (Flash 1.0) that reports 98.7% on the DF26 benchmark, a heatmap that pinpoints AI regions inside a photo, a lower 256×256 input floor, and access via web, browser extension, and API. It reads images and video about as well as anything available — but the figures are company-reported, a single score is still an estimate, and it generates and publishes nothing.
Pangram Image 1.0, released November 4, 2026, is the production version of the AI image detector Pangram first shipped as a research preview earlier in the year. Pangram is a New York detection company founded by Stanford AI graduates Max Spero and Bradley Emi, best known for its text detector and for powering Substack's reader-facing AI scan. This release pushes it firmly into visual media: a flagship image model (Image 1.0) plus Pangram Image Flash 1.0, a smaller, faster model built for video.
The case for it is strong on paper. Pangram reports 99.96% accuracy for Image 1.0 on an evaluation set of 1,200 synthetic images — 100 real photos mirrored with 12 frontier generators — and says it holds up across academic benchmarks including Synthbuster, Mirage, and Community Forensics. For video, it reports Flash 1.0 at 98.7% on the DF26 benchmark, which it frames as a 29-point lead over the next-best detector. The genuinely new capabilities are a heatmap that localizes where AI content sits in an image (so it flags a composited AI element, not just a whole-image verdict) and a minimum input size dropped from 512×512 to 256×256 pixels, which makes it usable on the recompressed images that circulate on feeds.
The honest caveats are the usual two, sharpened by novelty. The headline accuracy numbers are generated on Pangram's own evaluation sets; they're plausible and the method is disclosed, but they aren't independent audits, and AI-visual detection is an arms race against generators that improve monthly. And like any detector, a single result is an estimate, not proof — the risk isn't the tool, it's over-trusting one score when the stakes are high.
This review scores Pangram Image 1.0 as a visual detector — what it detects, how well, and how you reach it — and ends with an honest note on where Kompozy fits, which is not as a detection rival but as the engine that produces the visuals a tool like this then checks.
Pangram Image 1.0 is an AI image detector that returns a likelihood a picture is AI-generated and outputs a heatmap showing which regions look synthetic, so it can flag an AI object or face dropped into an otherwise real photo rather than scoring the whole image as one number. Pangram says it's trained for robustness to compression, edits, blur, noise, and low resolution, and it accepts inputs down to 256×256 pixels (below 512, results carry a reduced-confidence disclaimer). Pangram Image Flash 1.0, released alongside it, is a smaller model built for video: it analyzes an uploaded clip frame by frame so you can see which parts are flagged. You reach it through Pangram's web app (image detection and video upload), a browser extension that flags AI content as you browse, and an API that covers both image and video. It is part of Pangram's broader detection platform rather than a standalone product. What it is not is a creation or publishing tool — it scores visuals; it produces, brands, and distributes none.
Pangram Image 1.0 fits anyone whose deliverable is a judgment about a visual: newsrooms and platforms checking whether a photo or clip is synthetic, brands and agencies verifying UGC or stock before they run it, marketplaces screening listing images, and developers who need image or video detection via API. It's also a reasonable pre-publish self-check for a creator who wants to know how an AI-assisted visual will read. It's a weak fit for anyone who arrived hoping a detector would help them make visuals — it won't; once you have a heatmap, the work of producing captioned video, on-brand carousels, and face-consistent images is entirely ahead of you, and Pangram does none of it.
| Dimension | Score | Why |
|---|---|---|
| Image detection accuracy | 4.5 / 5 | A reported 99.96% on its own 1,200-image set plus named academic benchmarks — strong, though self-generated rather than independently audited. |
| Video detection (Flash 1.0) | 4.2 / 5 | Reports 98.7% on DF26 with frame-by-frame analysis and a claimed 29-point lead; capable but new and company-benchmarked. |
| Heatmap / AI localization | 4.4 / 5 | Pinpointing where AI content sits in a photo — including composited elements — is genuinely more useful than a single yes/no. |
| Robustness to compression & low-res | 4.3 / 5 | A 256×256 floor and explicit training for compression, blur, and edits make it usable on real feed images, not just clean exports. |
| Transparency of method & benchmarks | 3.9 / 5 | The mirror-based eval method and benchmark names are disclosed, but headline figures run on Pangram's own sets and per-benchmark numbers live in charts. |
| Access & integrations | 4.4 / 5 | Web app, a browsing-time extension, and an API covering both image and video cover most verification workflows. |
| Reliability as a sole judgment | 3.6 / 5 | A score is an estimate, not proof — the main risk is over-trusting one result for a consequential call, and detection is an ongoing arms race. |
| Pricing & value | 3.8 / 5 | Folded into Pangram's broader plans; the announcement links a general pricing page but doesn't spell out visual-tier limits, so confirm live. |
Pangram Image 1.0 is part of Pangram's broader detection platform rather than a separately priced product, and the launch announcement links a general pricing page without spelling out the specific image- and video-scan limits of each tier. Pangram has historically offered a usable free tier, a low individual plan (around $20/month), and a higher professional plan plus usage-based API and custom enterprise pricing, metered by scan volume — but because visual detection is heavier than text and these caps change often, reconcile any figure against pangram.com before budgeting.
Judged as a detector, paying for accuracy and access is reasonable value, especially for teams that verify at volume through the API. The place to be careful is interpretation, not price: a result you pay for is still an estimate, and the cost of over-trusting it — a wrongly flagged photographer, a real clip called synthetic — doesn't show up on the invoice.
The comparison people sometimes try to draw — a visual detector versus a content engine like Kompozy — is a category error. Pangram prices detection volume; Kompozy prices visuals generated and published. One tells you how an image or clip reads; the other makes it. If your need is verification, Pangram's pricing is the relevant one. If your need is producing and shipping visual content, a detector's price sheet is answering a different question.
| Use case | Fit | Why |
|---|---|---|
| Checking whether a photo is AI-generated | Strong | This is the core job, and the heatmap makes it more useful than a single opaque verdict. |
| Detecting AI-generated or manipulated video | Strong | Flash 1.0 analyzes clips frame by frame and reports top results on video benchmarks like DF26. |
| Newsrooms and platforms verifying visual provenance | Strong | Provider-agnostic detection plus an API fit authenticity and moderation workflows. |
| Spotting AI elements composited into a real photo | Strong | Localization is built for exactly this — it flags the region, not just the file. |
| Self-checking your own AI-assisted visuals before posting | OK | A useful pre-publish signal, though a single score is an estimate and the figures are company-reported. |
| Making finished, on-brand visual content | Weak | Pangram generates nothing — it only scores images and clips that already exist. |
| Publishing visuals across platforms | Weak | There is no scheduler or publisher; detection stops at the result. |
Kompozy is not an AI detector, and pitching it as a Pangram Image 1.0 rival would be dishonest — they sit at opposite ends of the visual workflow. Pangram reads a finished image or clip and tells you how AI it looks. Kompozy makes the visual: it's a content generation and publishing engine that produces Gemini face-locked Persona Photos and Persona Infographics, HeyGen avatar video (Persona Shorts, Persona HeyGen), Clipped and Marketing Shorts, Quote Graphics, and brand-exact HyperFrames carousels — all governed by a Persona Brief for a consistent identity — then reframes each per platform and publishes across eight social platforms plus blog and email from one review pipeline.
The reason the two belong in the same conversation is what this release signals: visual provenance is becoming a standard check the way text detection already did, and the losing response is to feed generator-default output through edits hoping the heatmap comes back clean. The winning response, and the one Kompozy is built around, is to produce visuals in a genuine, consistent brand identity and disclose how you make them, so a detector's verdict is a footnote rather than a threat. Use Pangram Image 1.0 as an honest proofing gate on the visuals where it matters; use Kompozy to produce the images, video, and carousels — and everything around them — in the first place.
It's among the stronger AI-visual detectors available: Pangram reports 99.96% on its own 1,200-image evaluation set and says Pangram Image Flash 1.0 scores 98.7% on the DF26 video benchmark. Those are strong figures, but they're company-reported on Pangram's own sets rather than independent audits — treat any single result as an estimate, and remember AI-visual detection is a fast-moving arms race.
For visual verification, yes. If your job is checking whether an image or clip is AI-generated — as a newsroom, platform, brand, marketplace, or developer using the API — it's accurate, provider-agnostic, and the heatmap adds real value. It's not worth buying if what you actually need is to make visual content, because it generates and publishes nothing.
Yes, via Pangram Image Flash 1.0, a smaller model released alongside it that analyzes uploaded clips frame by frame so you can see which parts are flagged. Pangram reports it at 98.7% on the DF26 benchmark, which it describes as a 29-point lead over the next-best detector — a company-reported figure.
Image 1.0 outputs a heatmap that localizes where AI content appears in a picture, rather than returning only a whole-image yes/no. That lets it flag an AI object or face composited into an otherwise real photo, which is useful for both verification and for self-checking which element of your own visual reads as AI.
For visual detection: Hive AI and Reality Defender for platforms and enterprises, Sensity AI for fraud and authenticity, and Google SynthID Detector for watermarked Google-model media. If your real goal isn't detecting visuals but producing them, the tool you want is a content engine like Kompozy, which generates and publishes on-brand images, video, and carousels across platforms.
Better than the earlier version. The minimum input size dropped from 512×512 to 256×256 pixels, and Pangram says the model is trained for robustness to compression, edits, blur, and noise — though results below 512 pixels carry a reduced-confidence disclaimer, so very small or heavily degraded inputs are less certain.
It's part of Pangram's broader detection platform, and the launch announcement links a general pricing page without detailing the visual-scan limits per tier. Pangram has historically offered a free tier plus paid individual, professional, and enterprise plans metered by scan volume — confirm current image and video limits and prices on pangram.com before budgeting.
That's the natural pairing. Generate your visuals in Kompozy — face-locked Persona Photos, avatar video, HyperFrames carousels, all held to a Persona Brief — then, where it matters, scan a piece in Pangram and read the heatmap to see which element reads as AI, regenerate that asset, and publish with a clear disclosure. Kompozy makes the visual; Pangram checks it.
See Pangram Image 1.0 vs Kompozy comparison → · Get Started →