OriginScope examines pixel-generation evidence rather than relying on appearance alone. It combines camera-process statistics, synthetic-generation signals, content-aware routing, and metadata inspection into an interpretable origin report.
Private browser analysis · image stays on this device
Choose a JPG, PNG, or WebP. You can also paste an image from your clipboard.
Nothing is uploaded in Private browser mode.OriginScope reports probabilistic forensic evidence. It does not establish authorship, legal ownership, or whether a caption/claim about an image is true.
The model card is visible to the person using the detector. Stronger numbers are never silently borrowed by the lightweight browser engine.
Full forensic engine: 79.17% grouped balanced accuracy, 95.56% routing accuracy on 90 images / 30 matched groups. Portable browser engine: lower-capacity model for privacy and static hosting. Dataset: MLLM-Generated-Image-Detection-Dataset ↗. These are development results, not a claim of universal real-world accuracy.
Nine bundled examples work without a server. When your private GPU tunnel is connected, OriginScope can also draw random samples from the much larger benchmark copy stored on your GPU server.
The repository includes a downloader and group-safe experiment runner for the public multi-thousand-image dataset. Matching scene variants can be kept together during evaluation to reduce leakage.
OriginScope keeps unsigned EXIF separate from pixel inference and cryptographic provenance. Missing metadata does not mean AI; camera metadata does not prove authenticity.
The current release verifies image-origin evidence. The same architecture can later support capture-time provenance and auditable evidence chains without changing the principle: show what can be inferred, what can be cryptographically proven, and where uncertainty remains.
Inspect code and experiments on GitHub →