# 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. Fast Facts - **Flagship Models**: `Pangram Image 1.0` (high-precision image detector with spatial localization heatmaps) and `Pangram Image Flash 1.0` (high-throughput frame-by-frame video and feed detector), detailed in the [Pangram Image 1.0 technical release](https://www.pangram.com/blog/introducing-pangram-image-1-0). - **Frontier Image Mirror Accuracy**: `99.96%` across `1,200` synthetic mirror images generated by `12` frontier AI image models from `100` real base photographs. - **DF26 Video Deepfake Benchmark**: `Pangram Image Flash 1.0` achieved `98.7%` accuracy on `DF26`, outperforming the runner-up detector by `+29.0 percentage points`. - **Video Synthetic Mirror Scale**: Tested across `1,529` total videos (`129` real social clips paired with `1,400` image-to-video synthetic mirrors across `14` frontier video generators). - **Low-Resolution & Compression Resilience**: Minimum supported resolution lowered by `75%` in pixel area from `512x512` down to `256x256` (with confidence notices below 512px) alongside adversarial training against social media JPEG/MP4 compression, blur, and noise. ## 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. | Feature / Benchmark Metric | Pangram Image 1.0 (Flagship Spatial) | Pangram Image Flash 1.0 (Video & Feed) | Legacy Academic / C2PA Metadata Checkers | |---|---|---|---| | **Primary Modality** | High-res Images & Hybrid Inpainting | **Native Video Uploads & Social Feeds** | Metadata-only or Uncompressed Stills | | **12-Generator Synthetic Image Mirror** | **99.96% Accuracy (1,200 Mirrors)** | High-speed screening | Fails when EXIF/C2PA stripped | | **DF26 Talking-Head Video Benchmark** | Supported via frame sampling | **98.7% (+29 pts over next best)** | Near-random on modern video diffusion | | **Spatial Localization Output** | **Pixel-level AI Heatmap Overlay** | **Frame-by-Frame Timeline Breakdown** | Binary pass/fail only | | **Minimum Input Resolution** | **256x256 px** (down from 512x512) | **256x256 px** | Typically requires >= 512x512 uncropped | | **Adversarial Compression Robustness** | Trained on re-encoded, blurred, noisy inputs | Resistant to TikTok/Reels/X transcoding | Brittle under heavy H.264/WebP re-encoding | ## 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. Next Steps 1. **Require Spatial Heatmap Verification for Insurance & Marketplace Uploads**: Integrate `Pangram Image 1.0` via API into claims, e-commerce refund, and KYC pipelines where bad actors use localized AI inpainting to fabricate damage on real photos. 2. **Deploy Frame-by-Frame Video Scanning for Executive & Trust Safety**: Use `Pangram Image Flash 1.0` to scan inbound video verification clips and social media brand mentions, catching hybrid videos where only a 3-second audio/lip segment was synthetically altered. 3. **Treat <512x512 Detections as Triage Signals**: When evaluating heavily cropped `256x256` thumbnails, pair Pangram's reduced-confidence flag with secondary source provenance checks.