Very Frame AI
How it works
Image and video use separate backend pipelines. The video path samples bounded frames, checks available signals across time, and retains only structured results.
- 1
Upload and validate
The API accepts an image and checks its file signature, size, image decoding, and dimensions.
- 2
Inspect file and quality
The backend records file details and checks image resolution and basic quality signals.
- 3
Read metadata and provenance
The backend reads EXIF availability and checks whether C2PA can be verified in its environment.
- 4
Run configured detectors
The backend runs the configured Hugging Face image-classification models (two distinct models are required for a verdict); their scores are not calibrated probabilities.
- 5
Check robustness and regions
The backend runs transformed-image checks and experimental tile analysis.
- 6
Aggregate and explain
The verdict pipeline combines detector outputs, quality, and robustness into a cautious verdict and stored report.
Video analysis
The video API inspects container/codec metadata with ffprobe, detects scene changes with FFmpeg, combines uniform and scene-change sampling within a configurable frame budget, filters unusable frames, then reuses the configured image-classification detectors on sampled frames. The deterministic verdict requires multi-frame evidence and sufficient scene coverage; sparse or unstable evidence remains inconclusive. The result timeline seeks the in-tab local preview to evidence timestamps when that preview remains available.
What is not implemented
Face detection/tracking, audio authenticity, audio/video synchronization, and motion consistency models are unavailable. The optional language model can draft a short explanation from structured video evidence and an independent grounding review can reject it; the model does not choose the video tools or override the deterministic verdict. A VISION_MODEL call for image analysis can send that image to the configured provider. Model IDs and labels must be verified by the operator; the project does not provide a verified Nemotron model ID.