Pangram launches new AI image and video detection models

New models for synthetic media

Pangram has released two new computer-vision models designed to identify AI-generated content. The first, Pangram Image 1.0, focuses on detecting synthetic images, while Pangram Image Flash 1.0 is a smaller, faster model built to analyze video content frame by frame.

These models represent an evolution from a previous research preview that only accepted images larger than 512 by 512 pixels. The new version 1.0 expands capabilities to include deepfakes and lower-resolution images, which were previously difficult for the system to handle with confidence.

The release also introduces several new features for users and developers. These include native video uploads, a dedicated dashboard, and an API that supports both image and video detection. Pangram also integrated the technology into a Chrome extension to flag AI material on various websites.

How the detection works

To ensure the models are not simply memorizing specific content, Pangram used a synthetic-mirror testing method. They took 100 real photographs and used 12 different image generators to create 1,200 synthetic versions that closely matched the original prompts.

This approach tests whether the detector can distinguish between a real scene and its AI-generated counterpart. By using similar subjects and scenes, the researchers ensure the model identifies the underlying synthetic patterns rather than just associating certain topics with AI generation.

The Image 1.0 model also features a heatmap output. This tool is designed to localize specific synthetic elements, such as an AI-generated object that has been placed inside an otherwise authentic photograph. This helps users see exactly which parts of an image are suspected to be fake.

Performance and benchmarks

Pangram reports that Image 1.0 identifies 99.2% of AI-generated images while maintaining a 0.27% false-positive rate. In their specific synthetic-mirror tests, the model achieved a 99.96% accuracy rate on the company-created set of images.

For video detection, the Flash 1.0 model was evaluated against the DF26 benchmark. This benchmark consists of 271 real and 2,420 synthetic videos of people speaking in direct-to-camera recordings, official statements, and studio interviews.

Pangram states that Flash 1.0 scored 98.7% on the DF26 benchmark. This result is 29 points higher than the next-best detector, which the benchmark authors noted performed close to random chance. The model is specifically designed to be robust against modern generation techniques.

Technical improvements and access

The updated models are trained to recognize common transformations that often occur when images are shared online. This includes detection of compression, downscaling, blur, and noise, which are often used by actors to try and hide the presence of AI.

While the minimum input size has been lowered to 256 by 256 pixels, Pangram includes a disclaimer for smaller images. They note that images below 512 pixels will produce a reduced-confidence warning, acknowledging that smaller files are harder to verify accurately.

Access to these tools has also been significantly widened since the initial preview. While the image API was previously invitation-only, it is now available to a broader audience. The integration into the Chrome extension allows for automated flagging across social media, shopping, and rental sites.

Practical use cases

For everyday internet users, the Chrome extension provides a way to see AI flags in real-time while browsing. This allows people to quickly identify synthetic content on social media feeds or retail websites without needing to manually run checks.

For developers and businesses, the updated API provides a way to build detection into their own products. This is particularly useful for platforms that need to moderate content or verify the authenticity of user-uploaded videos and photos.

By combining these models with multiple access points, Pangram aims to make detection more accessible. Whether a user is viewing a browser feed or a developer is processing a video upload, the system provides a consistent way to identify synthetic material.