Pangram, the Gold Standard for AI Text Checks, Releases Image 1.0 Tool to Sniff Fake Pics and Videos
Pangram has released its multimedia detection tool, claiming 99.96% accuracy in identifying AI-generated images.
Deep dive
- Detects AI-generated images with a claimed 99.96% accuracy.
- Uses 'Pangram Image 1.0 Flash' to perform frame-by-frame analysis on videos with 98.7% accuracy.
- Available as a Chrome extension, web tool, and API.
- Features heatmap generation to visualize AI-altered regions of an image.
- Supports images down to 256x256 pixels, though with reduced confidence.
- Benchmarked against competitors like TruthScan and Sightengine.
Decoder
- DF26: A benchmark dataset and evaluation standard used to measure the performance of deepfake and AI-generated video detection models.
Original article
Pangram, one of the most widely used and trusted AI text detectors out there, is finally releasing its multimedia detection tools to the masses. Pangram Image 1.0, which was previously available as a research preview, is now widely available on the web, through the company’s API, and via its eponymous Chrome extension.
The company claims its image detector can identify AI-generated pictures with an astounding 99.96% accuracy on its own evaluation set. Pangram also notes that it has improved the model’s accuracy over the past few months, especially against edited images, low-resolution pictures, and pictures made by the most advanced AI models out there. As far as the competition goes, TruthScan achieves an accuracy of 96.33%, followed by AI or Not at 91.38% and Sightengine at 88.33%.
Our video detection is powered by Pangram Image 1.0 Flash, a lighter-weight model that makes frame-by-frame detection possible. Pangram Image 1.0 Flash achieves 98.57% accuracy on AI videos, vs. 84.2% by the next best competitor.
When it comes to videos, Pangram has a separate, faster model called Pangram Image Flash 1.0, which checks videos frame by frame.
The company claims it offers 98.7% accuracy at detecting AI-generated videos, a 29-point lead over the next-best detector on the DF26 benchmark. TruthScan falls far behind at 84.2% accuracy, while Sightengine drops further to 81.8%.
Pangram says it compared its image and video detectors against models developed by academic researchers, where it consistently outperformed them on benchmark tests.
Notably, the model has been trained to see through subtle changes such as compression, blurring, artificial noise, and other artifacts that usually fool AI detectors. It can also highlight which parts of an image are AI-generated with a heatmap.
Moreover, Pangram Image 1.0 can now check images as small as 256 x 256 pixels, down from the 512 x 512-pixel minimum at launch, though results below 512 pixels come with a reduced-confidence warning.
To test its video detector, Pangram took 129 real videos from social media platforms and created AI-generated versions of them using 14 of the latest AI video models.
In addition to bringing Pangram Image 1.0 to its Chrome extension, the company is also letting users upload videos on its website, so the model can analyze them and tell users whether they’re AI-generated. The Chrome extension, meanwhile, flags AI content as you scroll across social media, online shopping platforms, rental listings, and other websites with images.
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi