Skip to content

VMAF NEG (No Enhancement Gain)

VMAF NEG is a model variant that penalises encoder in-loop sharpening and similar enhancements that inflate standard VMAF scores without delivering a genuine improvement in perceived quality for the viewer.

Background

Standard VMAF scores the similarity between a reference and a distorted video. It was trained on human opinion scores for content encoded with typical compression artifacts. One consequence: some encoders use in-loop sharpening (deblocking, deringing, adaptive sharpening) to boost apparent edge contrast, which the VMAF feature set interprets as closer to the reference even when the enhancement is perceptually neutral or mildly annoying.

The NEG variant was trained to be resistant to such enhancements. "NEG" stands for No Enhancement Gain: a sharpening filter that does not improve perceptual quality produces no additional VMAF score. This makes NEG appropriate for evaluating the underlying codec RD efficiency rather than the combined efficiency of codec plus post-processing.

When to use NEG

Use NEG when:

  • Comparing two codecs where one or both apply in-loop sharpening that could inflate its standard VMAF score. NEG gives a fairer comparison of the underlying compression quality.
  • Comparing encoder configurations (presets, rate-control strategies) within one codec where some configurations enable adaptive sharpening. NEG makes the curves comparable.
  • Evaluating codecs against a visually-lossless anchor (e.g. target VMAF 95+). Sharpening-boosted scores can conceal coding artefacts that are visible in critical viewing; NEG surfaces them.

Example: vmaf-tune compare --src source.yuv --neg --target-vmafs 94,96,97,98

When NOT to use NEG

Do not use NEG for:

  • Production quality monitoring against absolute baselines. If your delivery specification says "VMAF >= 93 for all segments", the threshold was almost certainly calibrated against standard VMAF scores. Using NEG will produce lower scores on the same content and lead to false alarms or unnecessarily high bitrates.
  • Archival quality gating where the original reference is enhancement-free. NEG penalises all sharpening relative to the reference, including benign codec sharpness recovery that genuinely improves perceived quality versus a blurry encode.
  • Live monitoring dashboards. The NEG score is not directly comparable to operator-facing SLA thresholds that were set using standard VMAF.

Model files

Model string File Use case
vmaf_v0.6.1neg model/vmaf_v0.6.1neg.json HD (up to 1080p) content
vmaf_4k_v0.6.1neg model/vmaf_4k_v0.6.1neg.json UHD / 4K content

Both files are in-tree in the model/ directory. Float-precision variants (vmaf_float_v0.6.1neg.pkl, vmaf_float_v0.6.1neg.pkl.model) are also present for the Python training harness.

Using NEG in vmaf-tune

The --neg flag is available on the following vmaf-tune subcommands:

Subcommand Effect
recommend Routes the scorer to the NEG model
tune-per-shot Each per-shot bisect uses the NEG model
compare All codec bisects in the run use the NEG model
ladder The convex-hull sampler scores with the NEG model
corpus The Phase A grid sweep scores with the NEG model

When --neg is passed together with --vmaf-model vmaf_v0.6.1, the model is routed to vmaf_v0.6.1neg. When --vmaf-model vmaf_4k_v0.6.1 is set (or selected automatically for 4K content), the model routes to vmaf_4k_v0.6.1neg.

Example: codec comparison with NEG

vmaf-tune compare \
    --src reference.yuv \
    --width 1920 --height 1080 \
    --neg \
    --target-vmafs 94,96,97,98 \
    --encoders libx264,libx265,libsvtav1 \
    --format json \
    --output compare_neg.json

Example: per-shot tuning with NEG

vmaf-tune tune-per-shot \
    --src reference.mp4 \
    --encoder libx265 \
    --target-vmaf 93 \
    --neg

Example: bitrate ladder with NEG

vmaf-tune ladder \
    --src reference.yuv \
    --resolutions 1920x1080,1280x720,854x480 \
    --target-vmafs 95,90,85 \
    --encoder libsvtav1 \
    --neg \
    --format hls

Using NEG directly with the vmaf CLI

The libvmaf CLI accepts a --model version=vmaf_v0.6.1neg flag:

vmaf \
    --reference reference.yuv \
    --distorted distorted.yuv \
    --width 1920 --height 1080 \
    --pixel_format 420 --bitdepth 8 \
    --model version=vmaf_v0.6.1neg \
    --json --output scores.json

Score interpretation

NEG scores are calibrated on a 0–100 scale, the same as standard VMAF. However, NEG scores are systematically lower than standard VMAF on the same content when the encoder uses any form of sharpening. This is by design — NEG is measuring a different quantity (codec compression efficiency without enhancement gain).

Do not compare NEG scores directly against standard VMAF thresholds or published quality grades. Establish NEG-specific operating points from your encoder sweep data.

Technical references