Skip to content

Tiny AI

The Tiny AI surface ships small ONNX perceptual-quality models alongside classic VMAF SVM models.

  • Overview — architecture, capabilities, model lifecycle
  • Training — train custom models with vmaf-train
  • Inference — run models via the C API or CLI
  • Benchmarks — latency and accuracy numbers
  • Security — op allowlists, model validation, supply chain
  • Bisect model quality — binary-search a checkpoint timeline for the first quality regression (also wired as a nightly CI gate)
  • Training data — Netflix corpus path convention, --data-root loader API, and evaluation harness for fork-local training runs
  • PTQ across EPs — investigate int8 PLCC drop on CPU vs CUDA vs OpenVINO (Arc / CPU) Execution Providers
  • Conformal VQA — distribution-free prediction intervals on top of any vmaf-tune predictor (split-conformal + CV+, no new dependencies, ADR-0279 implementation surface)
  • MOS-corpus ingestion family — unified index for KonViD-1k, KonViD-150k, LSVQ, YouTube UGC, Waterloo IVC 4K-VQA, LIVE-VQC, CHUG UGC-HDR, and BVI-DVC corpora; common schema, quick-start commands, aggregation workflow, and KonViD MOS head v1 entry point
  • Signal-mix audit — table-only coverage, redundancy, complementary-intersection, and blind-spot reports for refreshed AI feature tables
  • External benchmark wrappers — wrapper-only comparison harness for fork predictors, x264-pVMAF, and DOVER-Mobile
  • Second-opinion features — join out-of-tree NR/MOS scorer outputs into refreshed feature tables before retraining
  • MOS label materializer — join subjective MOS labels onto refreshed feature tables before real MOS-head training
  • CHUG UGC-HDR ingestion — local-only CHUG manifest/video ingest path for HDR subjective-MOS experiments
  • KonViD MOS head v1 — 5 081-parameter MLP that maps canonical-6 features + saliency + shot-metadata to a scalar subjective MOS prediction in [1.0, 5.0] (ADR-0336)
  • Saliency per-block evaluation — block-level IoU for saliency masks at the same granularity ROI encoders consume
  • Saliency feature materializer — append saliency_mean / saliency_var columns to existing JSONL or parquet feature tables before predictor and MOS-head retrains