`vmaf_tiny_v1` — legacy mlp_small VMAF fusion regressor¶

Status — Superseded. vmaf_tiny_v1 is retained for LOSO-eval baselines and quantization-epsilon regression fixtures only. For production use, prefer vmaf_tiny_v2 or later. See ADR-0244.

vmaf_tiny_v1 is the original tiny MLP fusion head trained over the canonical-6 libvmaf features (adm2, vif_scale0..3, motion2). It carries mlp_small architecture (Linear 6→16→8→1) with a separately shipped StandardScaler file, matching the early Phase-1 export recipe before scaler-baking (ADR-0244) was introduced. It is the single-split LOSO baseline referenced in docs/ai/loso-eval.md, docs/ai/quant-eps.md, and docs/ai/quantization.md.

Shipped checkpoint¶

Field	Value
Model id	`vmaf_tiny_v1`
Location	`model/tiny/vmaf_tiny_v1.onnx`
Architecture	`mlp_small` — Linear(6, 16) → ReLU → Linear(16, 8) → ReLU → Linear(8, 1)
Trainable parameters	257
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 and its `.data` file
ONNX opset	17
License	BSD-2-Clause-Patent
Registry entry	`vmaf_tiny_v1` in `model/tiny/registry.json` (`"smoke": true`)
SHA-256	`d30201dfa8a0cb1d6d5bbe342b0f9049e40bf86e57b2e3b14cbfcade9231e7a6`

Note on external data. The ONNX file uses the ONNX external-data format; the weight tensor lives alongside the .onnx file as vmaf_tiny_v1.onnx.data. Both files must be present for the model to load. This was repaired in PR #296 / PR #174 after an external-data filename mismatch was found in the initial commit.

Training corpus¶

The v1 checkpoint was trained on the Netflix Public Dataset only (9 reference sources × multiple encodings — local extract; not redistributed in-tree). The teacher score is vmaf_v0.6.1 (the classic SVM). This single-corpus baseline is what docs/ai/loso-eval.md references when reporting LOSO scores for the v1 architecture.

No KoNViD-1k or BVI-DVC rows were included at v1 training time. This is the key difference vs v2's 4-corpus union; v1 generalises less well to UGC content.

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 has not been re-evaluated on the full Phase-3 chain. The relevant LOSO figures appear in docs/ai/loso-eval.md as the single-split v1 baseline; they are not re-quoted here to avoid drift. For current production PLCC / SROCC / RMSE numbers, see vmaf_tiny_v2.md.

When to use this model¶

Use case	Recommendation
Production VMAF fusion	Use `vmaf_tiny_v2`
LOSO-eval historical baseline	`vmaf_tiny_v1` — required fixture
Quantization-epsilon regression	`vmaf_tiny_v1` — required fixture (ADR-0203)
Capacity comparison	Pair with `vmaf_tiny_v1_medium`

Runnable usage example¶

# Evaluate quality using the vmaf CLI with the 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.onnx \
    --json --output /tmp/vmaf_tiny_v1.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 — limited generalization to UGC content relative to the 4-corpus v2 checkpoint.
StandardScaler is not baked into the ONNX graph (unlike v2+), 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 as a fixture, not as a calibrated estimator.
Uses ONNX external-data format; both vmaf_tiny_v1.onnx and vmaf_tiny_v1.onnx.data must be co-located.
Superseded by vmaf_tiny_v2 for accuracy (PLCC improvement of 0.005–0.018 across the validation chain). Do not use v1 in new pipelines.

vmaf_tiny_v1 — legacy mlp_small VMAF fusion regressor¶