Pangram Image 1.0 release notes

Overview

Pangram has officially released Pangram Image 1.0 and Pangram Image Flash 1.0, marking a significant update to its AI detection capabilities. These models are designed to identify AI-generated content in both static images and video, offering features like heatmap-based localization, robustness against common image transformations, and frame-by-frame video analysis.

Pangram Image 1.0 for Still Images

Pangram Image 1.0 serves as the company's flagship model for image detection. It is built to maintain high accuracy even when images have undergone compression, downscaling, or malicious attempts to introduce noise or blur.

Key capabilities include:

  • High Precision: The model achieved 99.96% accuracy on a test set of 1,200 synthetic images generated by 12 different frontier AI models.
  • Localization: Users can view heatmaps that highlight the specific regions of an image identified as AI-generated.
  • Lower Resolution Support: The minimum input size has been reduced from 512x512 to 256x256, allowing for more flexible detection, albeit with a confidence disclaimer for lower resolutions.

Pangram Image Flash 1.0 for Video

Pangram Image Flash 1.0 is a specialized, high-speed model optimized for video detection. It addresses the difficulty of distinguishing between real and AI-generated video, particularly in content featuring human speech.

  • Performance: In tests against the DF26 academic benchmark, the model achieved 98.7% accuracy, representing a 29-point improvement over previous industry standards.
  • Methodology: The model performs frame-by-frame analysis to pinpoint AI-generated segments. To ensure fair evaluation, Pangram tested the model by creating "synthetic mirrors"—generating AI versions of real videos using the same source frames and descriptive prompts to minimize bias.

Tools and Integration

To support broader adoption, Pangram has introduced:

  • Chrome Extension: A browser-based tool that flags AI-generated content in social media feeds and online listings.
  • API Support: Updated endpoints now allow developers to integrate both image and video detection into their own applications.

Community Feedback and Discussion

The release has sparked discussion on Hacker News regarding the effectiveness and long-term viability of detection tools.

  • Robustness: Some users reported success using the tool on images captured from screens, suggesting the model picks up on persistent artifacts rather than just metadata. One user noted, "Taking a photo of an AI generated image on my grubby laptop screen with my phone, moire and all, still got detected."
  • The Arms Race: A common point of debate is whether detection systems create a feedback loop that aids in training better generators. As one commenter noted, "Any attempt to build AI generated content detection systems is flawed, since the outputs of such a system may be used to train an even better fake data generator."
  • Use Cases: While some users remain skeptical of AI detection for text, there is stronger consensus that image and video detection serves a critical function in verifying visual evidence. Users highlighted the importance of this technology for platforms like Instagram and WhatsApp to mitigate the spread of synthetic media.

Sources