GENERAITR Consistency
Headless drift detection for AI-generated media — spatial, temporal, and identity consistency, measured with the same metrics the video-generation research community uses to evaluate models.
The problem
Generative models drift: a character's face slowly morphs across a video, a product's shape shifts between "identical" variants, a mascot loses its identity across a campaign's worth of images. Each frame can look fine alone — the failure only shows up across outputs, which is exactly what a human skimming thumbnails won't catch, especially at the volume automated pipelines produce.
What it does
GENERAITR Consistency is a headless analysis service — MCP-first with a REST mirror — that takes generated media plus an optional reference and computes objective drift metrics across three dimensions: temporal (continuity within a video — morphing faces, warping logos, flicker), spatial (geometry and composition shifting within or across frames), and identity (whether separate generations of "the same" subject still are the same subject). It answers one question — has this drifted beyond what you'll accept — as a pass/fail gate or a full drift timeline. It detects and quantifies; it never fixes, regenerates, or judges aesthetics.
Why trust it
Built on the metrics the video-generation research community actually uses to evaluate models — RAFT optical-flow warping error, CLIP/DINOv2 consistency, InsightFace/ArcFace face embeddings, SSIM/LPIPS perceptual similarity — not ad-hoc heuristics. Deterministic: no LLM in the analysis core, so the same input always produces the same number. And honest by design: every result records exactly what was measured and what was skipped, so a dimension that couldn't be checked is reported as skipped, never silently folded into a passing score.
Current maturity
Working MVP: image sets and video are covered across all three drift dimensions (audio and native 3D are explicit non-goals for this stage). Threshold calibration is the one open, stated caveat — shipped thresholds are documented as unverified hypotheses pending calibration against real known-same/known-different data, and the system reports them as such rather than pretending otherwise. Built as its own standalone, deployable module from day one: it takes raw media in and works entirely independently of the rest of GENERAITR, so it can be adopted directly by an external pipeline.