Pangram releases Image 1.0 for AI-edit and video detection

Pangram released Image 1.0 and the smaller, faster Image Flash 1.0 on November 4 for image and video detection. The release adds heatmaps indicating areas believed to contain AI-generated edits and native video uploads with results across a timeline.

The launch announcement lowers the minimum image size from 512×512 pixels to 256×256. The website’s upload instructions still listed a 512×512 minimum when checked, leaving a discrepancy with the announcement. Inputs smaller than 512×512 receive a reduced-confidence notice. These are capabilities described by Pangram; this report did not independently measure detection performance.

Pangram says its classifier examines pixels and can work without watermarks or attached metadata. The service separately displays C2PA provenance records and EXIF information, bringing recorded production history into the inspection alongside the model’s assessment.

The image API distinguishes those evidence types. Model-based results can include an AI-likelihood score, confidence, model version and patch-level heatmap. When validated C2PA AI provenance determines the verdict, the response identifies C2PA as the source and can leave those model fields empty. A provenance-based verdict consequently may contain no model score.

Video inspection samples up to three frames per second

The video API accepts MP4, MOV and WebM files, extracts frames and analyzes them with Flash. It samples three frames per second, or the original rate when lower. Results include sampled-frame timestamps and judgments, classified timeline segments and an overall label: AI, mixed, no AI detected or unsure.

Uploads are limited to one file of 500 MiB. Videos running at three frames per second or above can be analyzed for up to two minutes, with at most 360 sampled frames. Over-limit files fail instead of being automatically trimmed. The default charge is $0.05 per second of inspected source duration, rounded up to a whole second: 30 seconds costs $1.50. Only successful tasks are billed. Completed results remain available for 24 hours.

The website offers the first three image scans each day without registration. It supports a single image or batches of up to ten files, combining inspection of suspected edits with available provenance information. That daily free allowance is described for images.

The Individual subscription, billed monthly, costs $20 and includes 100 image scans per month. A seven-day free trial precedes payment. Subscription image allowances and video API charges use different units; the published plan does not establish an allowance of 100 videos or 100 video minutes.

For integrations, the image API accepts one image per task, with a decoded-file limit of 32 MiB, and returns a task identifier. After a connection timeout, that identifier still supports polling. A fresh submission can incur another charge. Returned version information records the model used, while minor model updates can occur within the same selectable series.

Python SDK 1.1.0 adds native image and video methods. It accepts file paths or in-memory content, with a filename including the extension for video bytes. Submission and waiting can run together or separately. The client retries result polling after temporary transport errors, while submission itself is never automatically repeated.