Pangram's Image Detector Posts 99.96 Percent. It Also Built the Test Set.

The week Pangram Labs shipped its image detector, two photographers on opposite sides of the world were arguing with strangers about whether their pictures were real.
In New Zealand, astrophotographer Brendan Larsen spent two hours on Mount Taranaki in September capturing 578 frames, then about 45 hours layering them in Photoshop, most of the work spent erasing stars so roughly 1,400 satellites and 14,000 trails would show. The image went up in mid-October, drew tens of millions of views, and a slice of the audience called it AI. He had filmed his process in anticipation and posted it anyway, which is where things stand now: making the picture is the easy half.
In England, a photographer named Dave Wilson looked at a storm photo of Whitehaven lighthouse posted by an account called Aerial Perspectives and said the flight conditions made it impossible. He ran the file through OpenAI's Verify service, which checks C2PA metadata and SynthID watermarks, and got a hit. Checking the rest of the portfolio, he found almost all of it flagged. The company blamed a former team member for the storm image, acknowledged it should not have been presented as photography, and denied the portfolio was synthetic, explaining that it adds its own watermark through ChatGPT, which is why SynthID shows up.
That last explanation is either a reasonable account of a workflow or an extremely convenient one, and no detector on the market can currently settle it.
What Pangram shipped
Pangram built its reputation on text detection. On the same week in early November it extended to images and video with two models.
Pangram Image 1.0 is the flagship. The pitch is detection that survives compression, small files and partial manipulation, which is the actual condition of most images circulating online. The minimum resolution has dropped to 256 by 256 pixels, though the company flags results as less reliable when an image's shorter edge falls under 512 pixels.
Image Flash 1.0 is the smaller model, and it works on video frame by frame.
The accuracy numbers are where this gets interesting. Pangram reports 99.96 percent for Image 1.0 on a dataset it built itself, by having 12 frontier image generators produce 1,200 counterparts to 100 real photographs. Flash 1.0 scored 98.7 percent on DF26, a research study Pangram cites, and beat several major detection services on its own test set of 1,529 videos.
The German outlet heise, which reported the launch, noted plainly that none of these figures have been independently verified.
Why the test set is the whole argument
A detector's accuracy claim is only as meaningful as the images it was measured against. Pangram's flagship number comes from a balanced set where every real photo has twelve synthetic twins, one per generator, which is a clean experimental design. It also means the model was graded on the same twelve tools the company chose to include, and the published figure says nothing about how it handles a thirteenth.
That isn't a knock on Pangram specifically. It's how essentially every detection vendor reports, and it's why the honest way to read a number like 99.96 percent is as a ceiling measured under favourable conditions rather than a field estimate.
There's a second reason to hold the figure loosely. The EU AI Act's transparency requirements have been phasing in since August 2026, which means disclosure is becoming a compliance obligation rather than a courtesy. Regulation creates demand for a number a buyer can point at. A vendor reporting its own number is the predictable outcome of that demand, and independent evaluation is the thing that hasn't caught up.
The 256-pixel floor matters more than the headline
Buried under the accuracy claim is the more useful engineering detail. Most detection systems want a decent image. Social platforms recompress, crop and downscale everything, so a detector that works on a 256-pixel thumbnail is doing something different from one that needs the original file.

The caveat underneath it is equally practical. Below a 512-pixel edge, Pangram says results get less reliable. So the model will attempt a verdict on a small crop and will also tell you when that verdict is shaky.
For anyone doing verification work, that combination is more valuable than a single accuracy figure. A detector that says "I don't know" on a heavily cropped screenshot is more useful than one that always answers.
The video side is the harder problem
Image detection gets the attention because that's where the arguments happen, but video is where the gap has been widest. A single generated clip is hundreds of frames, and a watermark applied at the pixel level has to survive re-encoding, cropping to a vertical format, and whatever a social platform does to it on upload.
Pangram's approach is to check frame by frame with Flash 1.0. That is cheaper than treating a clip as one object, and it produces a distribution rather than a single verdict, which is closer to how a human reviewer works: scrub, stop somewhere suspicious, look closer.
The nuance it introduces is that a partly generated video now has a measurable answer rather than a yes or no. Advertising and short-form production both rely on hybrid footage, real plates with generated inserts, and the question that matters for disclosure rules is what fraction of the frames were synthesised.
What detection can and cannot do
The Aerial Perspectives case shows the limit clearly. A watermark hit is strong evidence that a generative model touched the file. A clean result proves nothing at all, because detectors only recognise watermarks from partners they track, and an image made elsewhere comes back clean. Cult of Mac made the same point about Apple adding SynthID to Image Playground in iOS 27.2: the check is one-directional.
That asymmetry is why the practical workflow for a photo editor now has three moves rather than one. Check the provenance metadata through a service like OpenAI Verify or Google's About This Image panel, which surfaces C2PA data when the file carries it. Reverse-search to find the earliest appearance. Then look at the picture, because hands, text and jewellery still give away more than any classifier.
Larsen's case adds the part nobody has solved. A real image built from 578 stacked frames looks synthetic because the process produces something the eye has never seen. Detection models have no category for that, and neither do the people leaving comments.
Pangram's launch is a real step. Video detection in particular has been thin, and a frame-by-frame model at that price point fills a gap. The claim to watch over the next few months isn't the 99.96 percent. It's whether anyone outside Pangram publishes a number that agrees with it.
Related articles
Six Amateurs, 1,113 Episodes, a Thousand Yuan of Compute an Episode: How an AI Drama Got Hand-Built
The first ten seasons of Armed Into Song run to roughly a thousand episodes and have passed ten billion cumulative views, with six core creators and not one of them film-school trained. Compute runs 1,000–2,000 yuan an episode; a normal shot takes three rolls, a battle twenty-plus, and anything that looks flat is graded a sixty and thrown out.
One server, one AI streamer: inside the code that runs a nightly livestream
open-yoki is the production code behind an AI host who streams nightly on Binance Square, with the persona, voice and money features stripped out. Headless Chrome draws the frame, ffmpeg pushes RTMP, a folder of Markdown decides what she feels and what she posts.
Pinky lets an agent drive your phone, and holds the Send button until you click Allow
Pinky connects any MCP agent to a real iPhone or Android phone and enforces approval in its own gateway rather than in a prompt. Banking apps are invisible, password fields are off limits, and on Android 13 the agent gets a hidden second screen.
ByteDance's SeedRealtime takes two frames a second and decides when to speak
Doubao-Seed-Realtime-1.0 folds camera frames, screen frames and speech into one full-duplex session, at no more than two frames per second. The interesting decisions are the frame rate, the proactive speaking, and the audit reply that ships switched on.