`vmaf_tiny_v1_medium` — legacy mlp_medium VMAF fusion regressor¶

Status — Superseded. vmaf_tiny_v1_medium is retained for LOSO-eval baselines alongside vmaf_tiny_v1. For production use, prefer vmaf_tiny_v3 (the mlp_medium successor) or later. See ADR-0244.

vmaf_tiny_v1_medium is the medium-capacity sibling of vmaf_tiny_v1, trained under the same Phase-1 recipe over the canonical-6 features (adm2, vif_scale0..3, motion2). It uses the mlp_medium architecture of the Phase-1 recipe (Linear 6→64→32→1 as shipped) without a baked-in StandardScaler. Both v1 variants are the single-split LOSO baselines referenced in docs/ai/loso-eval.md — v1 covers the mlp_small branch and v1_medium covers the mlp_medium branch for the capacity comparison. vmaf_tiny_v3 is the production mlp_medium-capacity checkpoint; it supersedes v1_medium in all production contexts.

Shipped checkpoint¶

Field	Value
Model id	`vmaf_tiny_v1_medium`
Location	`model/tiny/vmaf_tiny_v1_medium.onnx`
Architecture	`mlp_medium` — Linear(6, 64) → ReLU → Linear(64, 32) → ReLU → Linear(32, 1), read from the ONNX initialisers
Trainable parameters	2 561
Input	`input` — float32 `[batch, 6]`, dynamic batch
Feature order	`adm2, vif_scale0, vif_scale1, vif_scale2, vif_scale3, motion2`
Output	`score` — float32 `[batch, 1]`
Ops	`Gemm`, `Relu` (no scaler in the graph)
Sidecar	none ships; only the ONNX file
ONNX opset	17
License	BSD-2-Clause-Patent
Registry entry	`vmaf_tiny_v1_medium` in `model/tiny/registry.json` (`"smoke": true`)
SHA-256	`97f6116b44913f4076170a2f0cb78042db85aac7c56d432e14d9fe138ab952b7`

Training corpus¶

Identical to vmaf_tiny_v1 — Netflix Public Dataset only (9 reference sources; teacher score vmaf_v0.6.1). No KoNViD-1k or BVI-DVC rows. docs/ai/loso-eval.md references both vmaf_tiny_v1*.onnx as single-split LOSO baselines for the Phase-1 capacity sweep (small vs medium).

Training data terms¶

This model was trained on the Netflix Public Dataset. The terms below are quoted as each source states them (read 2026-10-04); the dataset terms list where each comes from and which models it trained.

Netflix Public Dataset: https://github.com/Netflix/vmaf/blob/0fb4152418d0351901e9c5fd2d30668dced89cdb/resource/doc/datasets.md

We provide a dataset publicly available to the community for training, testing and verification of results purposes.

(please request for access and we will grant it)

The stated purpose includes training; the page states no other terms.

Reading. The fork ships these weights under BSD-2-Clause-Patent: they are fitted parameters that cannot reproduce a clip, an image or a label, and no dataset file is redistributed. That is the fork's reading, not a permission from the dataset's authors. Where a dataset limits its use to research and that limit binds the weights where you use them, treat the model as research-only.

Retrain. RC9 retrains this model on data cleared for redistribution (T-TINY-AI-RETRAIN-CLEARED-DATA-2026-10-04 in state; ADR-1490, ADR-1570).

Validation¶

v1_medium has not been re-evaluated on the Phase-3 chain. For current production PLCC / SROCC / RMSE numbers at mlp_medium capacity, see vmaf_tiny_v3.md (trained on the 4-corpus union with baked StandardScaler, ADR-0275 PTQ sidecar).

When to use this model¶

Use case	Recommendation
Production VMAF fusion (medium capacity)	Use `vmaf_tiny_v3`
LOSO capacity baseline (`mlp_medium`)	`vmaf_tiny_v1_medium` — required fixture
Accuracy comparison vs `mlp_small`	Pair with `vmaf_tiny_v1`

Runnable usage example¶

# Evaluate quality using the vmaf CLI with the medium-capacity tiny model:
vmaf \
    --reference python/test/resource/yuv/src01_hrc00_576x324.yuv \
    --distorted python/test/resource/yuv/src01_hrc01_576x324.yuv \
    --width 576 --height 324 --pixel_format 420 --bitdepth 8 \
    --tiny-model model/tiny/vmaf_tiny_v1_medium.onnx \
    --json --output /tmp/vmaf_tiny_v1_medium.json

The tiny model is attached alongside the classic model and its score is added under the feature name vmaf_tiny_model (there is no sidecar name).

Input features

Loading the model makes the run compute its input features (adm2, vif_scale0..3, motion2, with default options), and the model scores every frame once the run is flushed. A frame without one of them fails the run with a message naming it; no input is read as 0.0 (ADR-1520).

Known limitations¶

Trained on Netflix Public Dataset only — same limited UGC coverage as vmaf_tiny_v1; vmaf_tiny_v3 uses the 4-corpus union.
StandardScaler is not baked into the ONNX graph (Phase-1 recipe), and no sidecar carrying the scaler statistics ships. libvmaf therefore feeds the raw features, so scores from --tiny-model are not scaler-corrected: treat v1_medium as a fixture, not as a calibrated estimator.
Superseded by vmaf_tiny_v3 for both accuracy and corpus coverage. Do not use v1_medium in new pipelines.

vmaf_tiny_v1_medium — legacy mlp_medium VMAF fusion regressor¶