VMAF v1 Models — Research Digest¶

Date: 2026-07-01 Author: VMAFx maintainers Upstream: Netflix/vmaf v3.2.0 (2026-06-20) ADR: ADR-1122 Sources: Netflix TechBlog "VMAF v1: Good Is Not Good Enough" (2026-06); Netflix/vmaf resource/doc/models_v1.md; model/vmaf_v1.0.16/*.json; libvmaf/src/model.c (built-in model table).

Summary¶

Netflix shipped a new generation of VMAF models — VMAF v1 — in libvmaf v3.2.0 (June 2026). v1 is not a new fusion architecture: it is still a libsvm ν-SVR (model_type: LIBSVMNUSVR) over quality-aware features. What changed is the feature set and the model calibration for viewing conditions. The headline change is that VIF is removed and CAMBI (banding) + a chroma feature are promoted to core features. Netflix also recommends measuring at 10-bit precision even for 8-bit SDR content, adds a 4K model with a [0, 110] score range, and provides high-frame-rate (_hfr) variants.

This digest records the exact model set, the feature delta, and the concrete implications for the fork (which already implements most of the v1 features on GPU). It backs ADR-1122.

The v1 model set (model/vmaf_v1.0.16/)¶

All models are libsvm ν-SVR JSONs, built into libvmaf (model.c exposes src_vmaf_v1_0_16_*_json), selectable via --model version=… or --model path=…json.

Plus _hfr variants under model/vmaf_v1.0.16_hfr/ (built-in as src_vmaf_v1_0_16_hfr_3d0h_json, …_hfr_3d0h_2160_json) calibrated for ~50/60 fps: a five-frame temporal motion window (differencing over frames i−2, i, i+2) with moving-average smoothing, correcting v0's under-prediction at high frame rates.

Feature delta (the core change)¶

The v1 standard model's feature_names (from vmaf_v1.0.16_3d0h.json):

Cambi_feature_cambi_score
Speed_chroma_feature_speed_chroma_uv_score
VMAF_integer_feature_adm3_score
VMAF_integer_feature_motion3_score

• VIF removed. Netflix found VIF is the most expensive feature and, after refining the others, dropped it without accuracy loss — so v1 is faster.
• CAMBI in the fusion. Banding (previously a standalone metric) is now a VMAF feature, addressing v0's blindness to banding.
• Chroma feature added (speed_chroma_uv) — v1 accounts for chroma distortion, which v0 ignored (luma-only).
• ADM → adm3, Motion → motion3 (specific variants; motion3 is the wider temporal window used more aggressively by the _hfr variants).

Key behavioral changes¶

1. 10-bit is the recommended precision. "VMAF v1 should ideally be applied at 10-bit precision for SDR" (models_v1.md) — even 8-bit encodes should be measured at 10-bit (preprocess both inputs) so CAMBI can resolve banding.
2. Scores can exceed 100. The 4K@3H model uses [0, 110] to quantify the 4K-over-1080p benefit at 3H. Any min(score, 100) assumption is now wrong.
3. CAMBI encode-side params. --model …:cambi.enc_width=W:cambi.enc_height=H: cambi.enc_bitdepth=B passes the pre-upscale encode geometry to the CAMBI feature; merged into the model's CAMBI options (no separate CAMBI instance). Falls back to input W/H/bitdepth when unset.
4. Selection unchanged. Same --model version=… / path=… mechanism as v0; v1 and v0 coexist (v0 preserved under models_v0.md).

Performance & accuracy (Netflix's claims)¶

• v1 matches or outperforms v0 on most datasets; notable gains on WATERLOO IVC 4K, Netflix Screen-Size Crowdsourcing, and datasets with chroma/banding artifacts and phone viewing.
• Faster than v0 (VIF — the compute bottleneck — is gone).
• Future: a detailed technical paper and an HDR v1 model are planned.

Implications for the fork (VMAFx)¶

The fork already implements almost every v1 feature on GPU/SIMD:

Net: v1 is a medium port, not a rewrite. The load-bearing work:

1. Port + build-in the 6 v1 JSONs (/add-model), matching Netflix's built-in table.
2. Fix SYCL CAMBI + speed_chroma — these become core features for every v1 score. speed_chroma_sycl currently needs fp64 (SIGABRT on Arc A380, ADR-0220 territory); cambi_sycl parity fails. These are v1 blockers on SYCL.
3. Handle the [0, 110] range — audit CLI/filters/--precision and any clamp-to-100. The FFmpeg libvmaf* filters and vf_libvmaf patches must not truncate.
4. VIF is now optional for v1 — keep for v0 compatibility, but the v1 path skips it (potential GPU speedup: the fork's whole VIF kernel stack is off the v1 critical path).
5. 10-bit alignment. v1's 10-bit recommendation directly validates the zero-copy P010 path (ADR-1121, PR #1081) — the fork is already positioned to measure v1 at the recommended precision, unlike 8-bit tooling (e.g. FFMetrics).
6. Golden gate. The 3 Netflix CPU golden pairs are v0.6.1 — do not touch (CLAUDE.md rule 1). v1 needs new golden references in separate tests.
7. Upstream sync. Netflix/vmaf v3.2.0 is the port target (/sync-upstream / /port-upstream-commit) — large: new models + VIF removal from fusion + CAMBI/chroma integration + CAMBI enc-params.

Open questions for the port¶

• Does adm3/motion3 in the fork bit-match upstream v1's feature definitions, or did upstream change the integer kernels for v1? (Verify against v3.2.0.)
• CAMBI enc-param plumbing through the FFmpeg filter AVOptions (ffmpeg-patch impact under CLAUDE.md rule 14).
• HDR v1 timing (Netflix "planned") — defer or scaffold.