Pangram Releases Image 1.0 and Flash 1.0 AI Detectors

Updated on November 4, 2026

Human reviewers can no longer reliably distinguish synthetic talking-head videos or localized AI image edits from authentic camera footage, and first-generation pixel detectors routinely fail when social media platforms downscale or recompress uploads. On November 4, 2026, AI detection lab Pangram released two production multimodal detection models—Pangram Image 1.0 and Pangram Image Flash 1.0—reporting 99.96% accuracy on frontier synthetic image mirrors and 98.7% accuracy on the academic DF26 deepfake video benchmark.

Authored by Pangram researcher Isaac Kasahara, the release graduates Pangram’s visual detection stack from its July 2026 research preview into a full API and Chrome Extension suite capable of both spatial heatmap localization (pinpointing the exact pixels where an AI object was inpainted into an otherwise authentic photograph) and frame-by-frame native video timeline analysis.

To avoid the dataset bias that plagues academic detectors—where models inadvertently learn to classify studio lighting or stock-photo subjects as “AI” rather than detecting generative artifacts—Pangram constructed a rigorous synthetic mirror evaluation. For images, the team took 100 real photographs and generated 1,200 synthetic twins across 12 frontier image generators using detailed scene-matched prompts; Pangram Image 1.0 achieved 99.96% accuracy across the paired dataset.

For video, Pangram took 129 real social media videos, seeded 14 different frontier AI video generators (100 videos per generator) with the actual first frame of each real clip plus a VLM-generated action description, and evaluated detectors across all 1,529 videos. On the academic DF26 talking-head video benchmark, Pangram Image Flash 1.0 scored 98.7%—a 29-percentage-point leap over the next-best competing detector.

Pangram Image 1.0 vs. Image Flash 1.0: Capabilities & Benchmark Matrix

Why did Pangram split its visual detection stack into two distinct models instead of a single monolithic checkpoint? Inspecting a single high-resolution insurance claim photo for localized inpainting requires dense spatial attention, whereas scanning a 60-second 30fps social media video requires evaluating dozens of frames per second at low latency.

Why First-Frame Seeded Video Mirrors Expose Weak Detectors

A common failure mode in enterprise deepfake detection occurs when a malicious actor takes a real 1-second video frame of an executive or property, feeds that authentic frame into an image-to-video diffusion model (Runway, Sora, Kling, Veo, or Hailuo), and generates a synthetic continuation. Detectors that rely on background texture anomalies or static lighting consistency get fooled because the scene composition, lighting, and subject identity are inherited directly from a real camera sensor.

By training and evaluating Pangram Image Flash 1.0 specifically on 1,400 first-frame-seeded video mirrors across 14 video models—as well as academic benchmarks AIGVDBench, DF26, and Ivy-Fake—the model learns to track temporal micro-inconsistencies across consecutive frames rather than relying on static background cues.