Pangram 发布 Image 1.0 与 Image Flash 1.0 两款 AI 图像视频检测模型

Pangram announced two computer-vision models on November 4, 2026: Pangram Image 1.0 for detecting AI-generated images and the smaller, faster Pangram Image Flash 1.0 for analyzing AI video frame by frame. The release also introduces heatmap output, native video uploads, Chrome extension integration and API support for images and video.

Image 1.0’s heatmap is intended to localize synthetic material within an image, such as an AI-generated element inside an otherwise real photograph. In an official launch post on X, Pangram said the model identifies 99.2% of AI-generated images while maintaining a 0.27% false-positive rate.

Paired Image and Video Tests

For its image evaluation, Pangram created synthetic mirrors of 100 real photographs with 12 image generators. Prompts closely matched each original photo, producing 1,200 synthetic images; Pangram reports 99.96% accuracy for Image 1.0 on that company-created set. The figure describes performance on this targeted pairing of real and synthetic images, rather than a universal accuracy rate for every image encountered online.

The video comparison applied a similar synthetic-mirror principle. Pangram started with 129 real social-media videos and used 14 video generators to create 100 synthetic videos per generator, sampled from the real set, for 1,529 videos in total. Each generator received the real video’s first frame and a vision-language model’s description of the action. Because each real video and its synthetic counterpart share a scene and subject, the design tests whether a detector can distinguish similar clips without rewarding a shortcut that associates particular content with AI.

The DF26 Video Benchmark

Pangram evaluated Flash 1.0 with DF26, a benchmark described in a preprint submitted on September 7, 2026. The benchmark contains 271 real and 2,420 synthetic videos produced by seven modern video models, focused on one person speaking in direct-to-camera recordings, official statements and studio interviews. Its authors report that both people and state-of-the-art deepfake detectors performed close to random chance, and argue for evaluations that explicitly test robustness to distribution shifts from modern generators.

Pangram says Image Flash 1.0 scores 98.7% on DF26, 29 points above the next-best detector. Native video uploads put that model into the dashboard: Flash analyzes each frame and identifies which portions of an uploaded video return an AI detection.

Deployment Beyond the Research Preview

The July 29, 2026 research preview accepted images only at 512 by 512 pixels or larger and listed deepfakes and face swaps as cases it could not confidently detect. Pangram says version 1.0 improves performance on deepfakes, low-resolution images and frontier generation models. The minimum input is now 256 by 256 pixels, but images below 512 pixels produce a reduced-confidence disclaimer.

Pangram says it trains the new model to detect transformations such as compression, downscaling, blur and noise, whether introduced by social platforms or by actors attempting to evade detection. Access has also widened since the preview, when image API access was invitation-only and dashboard video input and Chrome integration were listed as future work. Image 1.0 is integrated into the Chrome extension to flag AI-generated material across social media, shopping and rental sites, while the updated API supports both image and video detection.

The release’s practical significance is the combination of the new models with those access points: detection can run in a browser feed, on an uploaded video or through an API integration. The reduced-confidence disclaimer for smaller images remains the key qualification, showing that broader input support and wider distribution do not amount to uniform confidence across every image.