{"id":1243,"date":"2026-08-06T13:34:11","date_gmt":"2026-08-06T06:34:11","guid":{"rendered":"https:\/\/liveapi.com\/blog\/vmaf\/"},"modified":"2026-08-06T13:34:36","modified_gmt":"2026-08-06T06:34:36","slug":"vmaf","status":"publish","type":"post","link":"https:\/\/liveapi.com\/blog\/vmaf\/","title":{"rendered":"What Is VMAF? How Netflix’s Video Quality Metric Works and How to Use It"},"content":{"rendered":"Reading Time: <\/span> 13<\/span> minutes<\/span><\/span>

VMAF is how streaming engineers answer a deceptively hard question: does this encode actually look good? Netflix open-sourced the metric in June 2016, and it now predicts human viewer ratings on a 0-100 scale with a correlation of roughly 0.90-0.95.<\/p>\n

That’s far better than the pixel-difference math the industry relied on for decades.<\/p>\n

In June 2026, Netflix shipped VMAF v1, the first major model overhaul since v0.6.1. It changes which artifacts the metric can see, which defaults you get, and which model files you should be pointing at. If your pipeline still references vmaf_v0.6.1<\/code>, some of what follows will be news.<\/p>\n

What Is VMAF?<\/h2>\n

VMAF (Video Multimethod Assessment Fusion) is a full-reference video quality metric that predicts subjective viewer ratings by fusing several image-quality features through a trained machine learning model, producing a score from 0 to 100 for every frame.<\/p>\n

“Full-reference” is the part that trips people up. VMAF needs both files: the pristine source and the compressed output.<\/p>\n

It can’t look at a single video and tell you how good it is. It tells you how much quality the encode lost relative to what went in, which is exactly the question you’re asking when you tune a video codec<\/a> or build a bitrate ladder.<\/p>\n

The metric came out of a collaboration between Netflix, the University of Southern California, Nantes Universit\u00e9’s IPI\/LS2N lab, and the LIVE lab at the University of Texas at Austin. Netflix released it under a BSD+Patent license, so anyone can run it, retrain it, or ship it inside a product. The work earned a Technology and Engineering Emmy in 2021.<\/p>\n

VMAF exists because the old metrics lied. Two encodes could post identical PSNR numbers while one looked visibly worse to a person watching it. Netflix needed a number it could trust across thousands of titles and dozens of encoding decisions, so it built one from features that track how human vision actually fails.<\/p>\n

Here’s what the scale means in practice:<\/p>\n

\n\n\n\n\n\n\n\n\n\n\n
VMAF score<\/th>\nPerceived quality<\/th>\nTypical use<\/th>\n<\/tr>\n<\/thead>\n
90-100<\/td>\nExcellent: near-transparent to the source<\/td>\nPremium tier, mezzanine QC<\/td>\n<\/tr>\n
80-90<\/td>\nGood: minor artifacts under scrutiny<\/td>\nTop rungs of most ABR ladders<\/td>\n<\/tr>\n
70-80<\/td>\nFair to good: visible but acceptable<\/td>\nMid ladder rungs, mobile delivery<\/td>\n<\/tr>\n
50-70<\/td>\nFair: noticeable degradation<\/td>\nLow-bandwidth fallback rungs<\/td>\n<\/tr>\n
20-50<\/td>\nPoor: obvious artifacts<\/td>\nBelow most delivery thresholds<\/td>\n<\/tr>\n
0-20<\/td>\nBad: severe distortion<\/td>\nFailure state<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

VMAF vs PSNR vs SSIM<\/h2>\n

Three full-reference metrics dominate video engineering, and they measure genuinely different things. Picking the wrong one is how teams end up shipping encodes that score well and look bad.<\/p>\n

PSNR<\/strong> (peak signal-to-noise ratio) measures pixel-level error between two frames and reports it in decibels. It’s cheap to compute and it’s been the codec-comparison standard since the 1990s.<\/p>\n

It’s also close to blind to how people perceive video. A blurry frame and a blocky frame can post the same PSNR<\/a> while looking nothing alike to a viewer.<\/p>\n

SSIM<\/strong> (structural similarity index) improved on that by comparing luminance, contrast, and structure across local neighborhoods instead of individual pixels. MS-SSIM, the multi-scale variant, tracks perception better still. Both remain frame-by-frame measurements with no sense of motion.<\/p>\n

VMAF<\/strong> fuses multiple features, including a temporal one, through a model trained on actual human ratings. That training step is the difference. PSNR and SSIM are formulas someone derived; VMAF is a prediction fitted to what viewers said.<\/p>\n

\n\n\n\n\n\n\n\n\n\n\n\n
Attribute<\/th>\nPSNR<\/th>\nSSIM<\/th>\nVMAF<\/th>\n<\/tr>\n<\/thead>\n
What it measures<\/td>\nPixel error<\/td>\nStructural similarity<\/td>\nPredicted viewer opinion<\/td>\n<\/tr>\n
Scale<\/td>\nDecibels, logarithmic<\/td>\n0-1, logarithmic feel<\/td>\n0-100, roughly linear<\/td>\n<\/tr>\n
Temporal awareness<\/td>\nNone<\/td>\nNone<\/td>\nYes (motion feature)<\/td>\n<\/tr>\n
Trained on human ratings<\/td>\nNo<\/td>\nNo<\/td>\nYes<\/td>\n<\/tr>\n
Compute cost<\/td>\nVery low<\/td>\nLow<\/td>\nModerate to high<\/td>\n<\/tr>\n
Correlation with MOS<\/td>\nWeak<\/td>\nModerate<\/td>\nStrong (~0.90-0.95)<\/td>\n<\/tr>\n
Handles scaling artifacts<\/td>\nPoorly<\/td>\nPartially<\/td>\nYes<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

The linear scale matters more than it sounds. Moving PSNR from 38 dB to 40 dB tells you almost nothing about how much better the video looks. Moving VMAF from 85 to 93 maps to something a viewer would actually notice, which makes it usable as a target in automated encoding decisions and in quality of experience<\/a> reporting.<\/p>\n

None of this makes PSNR useless. It’s fast, it’s deterministic, and it’s a decent smoke test for “did this encode break.” Most production pipelines compute all three and use VMAF for the decisions that matter.<\/p>\n

How Does VMAF Work?<\/h2>\n

VMAF doesn’t compute one thing. It computes several elementary features per frame, then feeds them into a trained regressor that outputs a single number.<\/p>\n

    \n
  1. Align the two videos.<\/strong> Reference and distorted clips have to match in resolution, frame rate, and frame count. Mismatch here is the single most common source of nonsense scores.<\/li>\n
  2. Extract elementary features per frame.<\/strong> The v0 models use Visual Information Fidelity (VIF) at multiple spatial scales, the Detail Loss Metric (DLM), and Mean Co-Located Pixel Difference (MCPD) for motion. VMAF v1 drops VIF, keeps DLM and MCPD, and adds a banding detector and chroma features.<\/li>\n
  3. Feed the features into the fusion model.<\/strong> A support vector machine regressor, trained on subjective scores from human viewing sessions, maps the feature vector to a predicted opinion score.<\/li>\n
  4. Clamp to the score range.<\/strong> Each frame gets a value from 0 to 100, where 100 means indistinguishable from the reference.<\/li>\n
  5. Pool the per-frame scores.<\/strong> By default VMAF takes the arithmetic mean across every frame to produce the clip-level score you see reported.<\/li>\n<\/ol>\n

    The motion feature is what separates VMAF from its predecessors. MCPD measures how much the luminance channel changes between consecutive frames, which lets the model account for temporal masking: viewers tolerate far more compression artifacts in fast-moving content than in a slow pan across a gradient sky.<\/p>\n

    A frame-by-frame metric can’t make that distinction. That’s why per-frame metrics penalize action sequences that look fine and wave through static scenes that look terrible.<\/p>\n

    The per-frame output is also the part most teams ignore, and the clip average hides everything. A file averaging VMAF 92 with a 30-frame dip to 60 produces a visible glitch the headline number never reveals.<\/p>\n

    That kind of dip is a common cause of complaints about video encoding<\/a> quality that “tests fine.”<\/p>\n

    VMAF Models Explained: v0.6.1 vs VMAF v1<\/h2>\n

    VMAF isn’t one model. It’s a family, and the one you pick encodes an assumption about how your viewers watch. Point at the wrong file and your scores will be internally consistent and externally meaningless.<\/p>\n

    The v0.6.1 family (2016-2026)<\/h3>\n

    vmaf_v0.6.1<\/code> is the default model almost every tutorial and blog post references. It assumes a 1080p display viewed from three times the screen height, which is roughly a TV across a living room. Netflix added a phone model in 2017 that applies a polynomial adjustment for small screens held close, and a 4K model in 2018 trained for 4K<\/a> TVs at 1.5x display height.<\/p>\n

    There’s also vmaf_v0.6.1neg<\/code>, the no-enhancement-gain variant. Netflix published it after researchers showed that sharpening a video before encoding could inflate its VMAF score without improving how it looked. NEG caps the reward for enhancement, which makes it the honest choice for comparing encoders.<\/p>\n

    The VMAF v1 family (June 2026)<\/h3>\n

    VMAF v1 is the first structural rewrite in a decade, and it changes four things:<\/p>\n