--- title: "Video Quality Metrics: VMAF, PSNR and SSIM Explained" description: Understand VMAF, PSNR, and SSIM—key video quality metrics—and learn how PSNR vs SSIM compare for video assessment. image: https://blog.blazingcdn.com/hubfs/Gemini-Blog/image-Oct-25-2026-08-30-18-1503-AM.jpeg --- [![BlazingCDN](https://blog.blazingcdn.com/hubfs/Logo/blog-logo-w.png)](https://blog.blazingcdn.com?hsLang=en-us) [📘Learn ▾](https://blog.blazingcdn.com/cdn-learn?hsLang=en-us) [CDN Fundamentals](https://blog.blazingcdn.com/cdn-fundamentals?hsLang=en-us) [By Content Type](https://blog.blazingcdn.com/by-content-type?hsLang=en-us) [Advanced Concepts](https://blog.blazingcdn.com/advanced-concepts?hsLang=en-us) [Glossary](https://blog.blazingcdn.com/glossary?hsLang=en-us) [⚡ Web Performance](https://blog.blazingcdn.com/web-performance?hsLang=en-us) [🎬Video & Streaming ▾](https://blog.blazingcdn.com/video-streaming-cdn?hsLang=en-us) [🔴 Live Streaming](https://blog.blazingcdn.com/live-streaming?hsLang=en-us) [📺 VOD & OTT](https://blog.blazingcdn.com/vod-ott?hsLang=en-us) [💰 Bandwidth & Costs](https://blog.blazingcdn.com/bandwidth-costs?hsLang=en-us) [🏭 Other Industries ▾](https://blog.blazingcdn.com/cdn-industry-insights?hsLang=en-us) [📺 Media & Broadcasting](https://blog.blazingcdn.com/media-broadcasting?hsLang=en-us) [💾 Software & SaaS](https://blog.blazingcdn.com/software-saas?hsLang=en-us) [🏗️ DevOps & Cloud Infra](https://blog.blazingcdn.com/devops-cloud-infra?hsLang=en-us) [📚 AdTech & Advertising](https://blog.blazingcdn.com/adtech-advertising?hsLang=en-us) [🎮 Gaming & Esports](https://blog.blazingcdn.com/gaming-esports?hsLang=en-us) [📱 Mobile Apps & Developers](https://blog.blazingcdn.com/mobile-apps-developers?hsLang=en-us) [🤖 AI & Machine Learning](https://blog.blazingcdn.com/ai-machine-learning?hsLang=en-us) [🏛️ Enterprise & Corporate](https://blog.blazingcdn.com/enterprise-corporate?hsLang=en-us) [📚 E-Learning & EdTech](https://blog.blazingcdn.com/e-learning-edtech?hsLang=en-us) [🏟️ Sports & Live Events](https://blog.blazingcdn.com/sports-live-events?hsLang=en-us) [💰 Pricing & Costs ▾](https://blog.blazingcdn.com/cdn-pricing-and-cdn-costs?hsLang=en-us) [Provider Pricing](https://blog.blazingcdn.com/provider-pricing?hsLang=en-us) [Cost Optimization](https://blog.blazingcdn.com/cost-optimization?hsLang=en-us) [Decision Support](https://blog.blazingcdn.com/decision-support?hsLang=en-us) [⚡Compare ▾](https://blog.blazingcdn.com/cdn-comparison?hsLang=en-us) [Provider Comparisons](https://blog.blazingcdn.com/provider-comparisons?hsLang=en-us) [Strategy Comparisons](https://blog.blazingcdn.com/strategy-comparisons?hsLang=en-us) [Ratings & Benchmarks](https://blog.blazingcdn.com/cdn-ratings-and-benchmarks?hsLang=en-us) [📊 Benchmarks](https://blog.blazingcdn.com/cdn-ratings-and-benchmarks?hsLang=en-us) [🔒 Security ▾](https://blog.blazingcdn.com/cdn-security?hsLang=en-us) [Attack Protection](https://blog.blazingcdn.com/attack-protection?hsLang=en-us) [Encryption & Access](https://blog.blazingcdn.com/encryption-access?hsLang=en-us) [Video & DRM Security](https://blog.blazingcdn.com/video-drm-security?hsLang=en-us) [📁 Case Studies](https://blog.blazingcdn.com/case-studies?hsLang=en-us) [🔌 Integrations](https://blog.blazingcdn.com/integrations?hsLang=en-us) [🛠️ Tools](https://blog.blazingcdn.com/cdn-tools?hsLang=en-us) [Get Started](https://blazingcdn.com/sign-up-contact-form/) [Learn](https://blog.blazingcdn.com/en-us/tag/learn) [Video - Video & Streaming](https://blog.blazingcdn.com/en-us/tag/video-video-streaming) [Learn - CDN Fundamentals](https://blog.blazingcdn.com/en-us/tag/learn-cdn-fundamentals) [Benchmarks](https://blog.blazingcdn.com/en-us/tag/benchmarks) # Video Quality Metrics: VMAF, PSNR and SSIM Explained Vadym Borovskyi Oct 25, 2026, 9:32:22 AM ![](https://blog.blazingcdn.com/hubfs/Gemini-Blog/image-Oct-25-2026-08-30-18-1503-AM.jpeg) VMAF, PSNR and SSIM are the three video quality metrics most streaming teams run in encoding pipelines, and each answers a different question. PSNR measures pixel error in decibels, SSIM measures structural similarity on a 0 to 1 scale, and VMAF (Video Multimethod Assessment Fusion) predicts a viewer's opinion score on a 0 to 100 scale. For bitrate ladder decisions, VMAF is the metric to trust. A gap of about 6 VMAF points is the working threshold for a just-noticeable difference (JND). As of 2026, VMAF, open-sourced by Netflix in 2016, is the standard metric for per-title bitrate ladder tuning. A 1080p rendition scoring 93 to 95 on the default model is generally treated as visually transparent on a living-room TV, and adjacent rungs closer than about 6 VMAF points are candidates for removal. PSNR and SSIM remain useful as fast regression checks, not as ladder targets. ## **How VMAF, PSNR and SSIM score a frame** All three are full-reference metrics. Each compares a decoded, distorted frame against the pristine source frame, pixel-aligned and at the same resolution. They differ in what they consider an error and in how closely their output tracks a human viewer. ### PSNR: error energy in decibels Peak Signal-to-Noise Ratio is 10 times log10 of MAX squared divided by the mean squared error (MSE), where MAX is 255 for 8-bit video and 1023 for 10-bit. Most tools report it on luma (Y-PSNR) or as a 6:1:1 weighted YUV average. Lossy streaming encodes typically land between 30 and 50 dB. Above roughly 40 dB at 1080p, artifacts are hard to spot, and below 30 dB they are usually visible. PSNR is blind to where the error lands. Uniform low-level noise and a single blocky macroblock edge can produce the same MSE, while viewers forgive the first and notice the second immediately. ### SSIM: local structure, luminance and contrast The Structural Similarity Index, introduced in 2004, compares local windows (commonly 8x8 or an 11x11 Gaussian) on three terms: mean luminance, variance as contrast, and covariance as structure. The product is averaged across the frame. MS-SSIM repeats the comparison at several downscaled resolutions to approximate different viewing distances. SSIM's scale is compressed near the top. The perceptual gap between 0.98 and 0.99 can be larger than between 0.90 and 0.95, which is why many encoder teams convert it to decibels as minus 10 times log10 of (1 minus SSIM) before plotting curves. ### VMAF: a learned model of viewer opinion VMAF does not measure distortion directly. It fuses elementary features (Visual Information Fidelity at four scales, a Detail Loss Metric, and a luma motion feature) through a support vector regression trained on subjective scores from controlled viewing sessions. The output predicts how a viewer would rate the clip on a 0 to 100 scale. The model file matters as much as the score. The default 0.6.1 model assumes a 1080p display viewed at three times screen height, the 4K model assumes 1.5 times screen height, and a phone model scores the same encode higher because small screens hide detail loss. A NEG (no enhancement gain) variant exists to stop sharpening filters from inflating scores. ## **PSNR vs SSIM vs VMAF: which metric for which job** PSNR and SSIM are cheap, deterministic and fine for catching regressions between two builds of the same encoder. VMAF costs more compute but correlates far better with viewer ratings across resolutions and content types, which makes it the only one of the three suited to choosing ladder rungs and resolution switch points. | Metric | Range | Common "good" threshold at 1080p | Relative compute cost (estimate) | Best use | | --- | --- | --- | --- | --- | | PSNR | About 20 to 60 dB | 40 dB or higher | 1x | Encoder regression tests, codec research baselines | | SSIM / MS-SSIM | 0 to 1 | 0.95 or higher | 2x to 5x | Structural damage checks, tuning deblocking and AQ | | VMAF | 0 to 100 | 93 to 95 for the top rung | 10x or more | Ladder design, resolution switch points, A/B of encode settings | Use PSNR and SSIM to catch breakage cheaply, and spend the extra VMAF compute only on decisions that change what viewers see and what you pay in egress. ## **What VMAF scores mean in practice** A single VMAF number is a pooled average of per-frame scores, and the arithmetic mean hides the frames viewers complain about. A 60-minute title can average 94 while a 20-second dark, grainy scene drops to 75. Viewers remember that scene, not the average. Report at least three numbers per encode: the arithmetic mean, the harmonic mean (which penalizes low frames harder), and the 1st or 5th percentile frame score. A practical rule of thumb many teams apply: when the mean minus the 5th percentile exceeds about 10 points, the problem is a scene, not the rung, and the fix belongs in per-shot encoding or rate control rather than in a higher bitrate across the whole title. Interpret differences, not absolutes. Moving from 82 to 88 is roughly one JND and visible to most viewers in side-by-side tests. Moving from 96 to 98 typically costs 20 to 40 percent more bitrate (estimate, content dependent) for a change few viewers can see on a TV, and none on a phone. ## **How to tune a bitrate ladder with VMAF** The standard method is a convex hull search per title or per shot. You encode the same content at several resolutions and quality points, score every encode with VMAF at the display resolution, and keep only the encodes that sit on the upper envelope of the quality versus bitrate curve. 1. Encode each title at five or six resolutions (for example 1080p, 720p, 540p, 432p and 360p), each at six to eight constant-quality points. 2. Upscale every encode to the reference resolution with a fixed filter (bicubic is the common choice) before scoring, because the player upscales too and VMAF must judge what the viewer sees. 3. Plot VMAF against measured average bitrate for every encode, and keep the points that form the upper convex hull. 4. Place the top rung where the curve flattens, for example where an extra 1 Mbps buys less than 1 VMAF point. 5. Walk down the hull, keeping a rung roughly every 6 VMAF points or every 1.5x to 2x bitrate step, whichever comes first. 6. Choose the resolution for each rung from whichever resolution owns the hull at that bitrate, which is where switch points usually move compared with a fixed ladder. This blog's pruning heuristic: drop any rung whose VMAF is within 4 points of its upper neighbor and whose bitrate is within 1.4x of it. Players rarely gain stability from such a rung, and every extra rung adds encode time, storage and cache footprint. ### Worked calculation: what a tuned top rung saves in egress Suppose a fixed ladder ships its 1080p top rung at 8 Mbps, and the convex hull shows the same title reaches VMAF 95 at 5.8 Mbps. One viewing hour at 8 Mbps is 8,000,000 bits times 3,600 seconds divided by 8, or 3.6 GB. At 5.8 Mbps it is 2.61 GB. At 100,000 top-rung viewing hours a month, that is 360 TB against 261 TB: 99 TB of egress removed with no measurable quality change. Priced on BlazingCDN's progressive tiers, which run from $5 per TB down to $2.50 per TB at volume as of October 2026, 360 TB costs $1,325 a month and 261 TB costs $978.50, so the tuned ladder saves $346.50 a month on that one rung (current rates on the [**BlazingCDN CDN pricing page**](https://blazingcdn.com/pricing/)). On metered providers with higher per-GB list prices, the same 99 TB is worth proportionally more. ## **Verify this on your own stack** Build ffmpeg with libvmaf enabled and run the libvmaf filter with the distorted encode as the first input and the source as the second, after scaling the distorted stream to source resolution with bicubic. Enable the PSNR and SSIM features in the same pass and set the log format to JSON so every frame is recorded. - Check the pooled mean, harmonic mean and per-frame minimum in the JSON log, not just the console summary. - Confirm frame counts match between source and encode; a single dropped or duplicated frame shifts alignment and can cut VMAF by 20 points or more. - Confirm both inputs share pixel format and color range (limited versus full), since a range mismatch moves PSNR by several dB. ## **Trade-offs and failure modes of video quality metrics** VMAF can be gamed. Sharpening, contrast boosts and some denoisers raise the default model's score without improving what viewers see, which is why the NEG model exists. Report both when you evaluate pre-processing filters. Film grain is a known blind spot. AV1 film grain synthesis strips grain before encoding and re-adds a synthetic version at decode, so a full-reference metric sees large pixel differences and scores the encode lower even when viewers prefer it. Score grain-heavy titles with grain synthesis disabled, or rely on subjective review for them. HDR is outside the comfort zone of the standard models, which were trained on SDR viewing. Teams shipping PQ or HLG content should treat VMAF numbers as relative within a pipeline, not as absolute thresholds. PSNR has the opposite bias: it rewards blur, because smoothing reduces MSE, so a PSNR-only tuning loop drifts toward soft, plastic-looking encodes. Compute is the operational cost. Full VMAF on every encode of a large catalog is expensive, so many pipelines subsample every third or fifth frame, which usually keeps the pooled mean within about 1 point but can miss a short bad scene. ## **Who this fits and who it does not** Convex hull tuning with VMAF pays off for VOD catalogs with heavy, repeated viewing, where encoding cost is paid once and egress savings recur every month. It also fits sports and premium live events where a pre-tuned ladder per content class is feasible. It fits poorly for low-traffic libraries, where the compute spent on six resolutions times eight quality points outweighs any egress saving, and for user-generated content with unpredictable sources. In those cases a fixed ladder with constant-quality encoding and a PSNR or SSIM regression gate is usually the better trade. ## **FAQ: VMAF, PSNR and SSIM video quality metrics** ### What is a good VMAF score for streaming? A VMAF score of 93 to 95 on the default model is generally treated as visually transparent for a 1080p top rung on a TV. Lower rungs commonly sit between 60 and 85. Interpret scores relative to the model used: the phone model rates the same encode higher, and the 4K model assumes a closer viewing distance. ### Is PSNR or SSIM better for measuring video quality? SSIM usually tracks viewer perception better than PSNR because it compares local structure instead of raw pixel error. Both still miss temporal artifacts and correlate weaker with viewer ratings than VMAF. PSNR remains the codec research baseline, while SSIM is useful for catching blocking and structural damage during encoder tuning and regression tests. ### How many VMAF points is a noticeable difference? About 6 VMAF points is the commonly used just-noticeable-difference threshold for 1080p content on a TV. Differences of 2 points or less are rarely visible to viewers. Many teams use the 6-point spacing to place bitrate ladder rungs, which keeps each step visibly different while avoiding redundant renditions that add encode and storage cost. ### Why do I need to upscale before computing VMAF? VMAF compares frames pixel by pixel, so the distorted video must match the reference resolution. Upscaling a 720p rendition to 1080p with a fixed bicubic filter mimics what the player does on a 1080p screen. Changing the scaler between runs can move VMAF by several points, so keep the filter constant across every comparison. ## **Score your current ladder before buying more bandwidth** This week, take your five most-watched titles and score every rung of the current ladder with VMAF at display resolution, logging mean, harmonic mean and 5th percentile. Then encode the top rung at three lower bitrates and find the lowest one that stays within 2 VMAF points of today's score. Multiply the bitrate difference by your monthly top-rung viewing hours, as in the worked calculation above, and you have a defensible egress number for the next planning cycle. If those titles are delivered as packaged HLS or DASH segments, the [**BlazingCDN Video CDN for VOD libraries**](https://blazingcdn.com/video-cdn/) replicates the library inside the CDN and keeps the origin out of the delivery path, with a 14-day testing period on real production traffic. Share: [f](https://www.facebook.com/sharer/sharer.php?u=https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained) [in](https://www.linkedin.com/sharing/share-offsite/?url=https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained) [𝕏](https://twitter.com/intent/tweet?url=https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained&text=) [✉](mailto:?subject=%3Cspan%20id="hs_cos_wrapper_name"%20class="hs_cos_wrapper%20hs_cos_wrapper_meta_field%20hs_cos_wrapper_type_text"%20style=""%20data-hs-cos-general-type="meta_field"%20data-hs-cos-type="text"%20%3EVideo%20Quality%20Metrics:%20VMAF,%20PSNR%20and%20SSIM%20Explained%3C/span%3E&body=https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained) ![BlazingCDN](https://blog.blazingcdn.com/hs-fs/hubfs/Logo/blog-logo-w.png?height=24&name=blog-logo-w.png) *Heavy traffic.* Light bill. The CDN for video and large traffic Their monthly bill vs ours - 20 TBFastly $2,087**$92.50** - 50 TBCDN77 $990**$215** - 200 TBCloudFront $11,965**$765** Published list prices, Aug 2026 [Calculate your cost](https://blazingcdn.com/cdn-cost-calculator/?utm_source=blog&utm_medium=sidebar&utm_campaign=blog_sidebar&utm_content=compare_calc) ## Related posts [![](https://blog.blazingcdn.com/hubfs/Gemini-Blog/image-Oct-25-2026-08-00-30-7093-AM.jpeg)](https://blog.blazingcdn.com/en-us/what-are-cheap-cdn-providers?hsLang=en-us) Pricing - Pricing & Costs ### [Cheap CDN Providers 2026: Low-Cost Options That Still Perform](https://blog.blazingcdn.com/en-us/what-are-cheap-cdn-providers?hsLang=en-us) A cheap CDN in 2026 costs $415 to $1,000 for 100 TB of monthly delivery at published list prices, while hyperscaler ... Oct 25, 2026, 9:02:55 AM [Read more](https://blog.blazingcdn.com/en-us/what-are-cheap-cdn-providers?hsLang=en-us) [![](https://blog.blazingcdn.com/hubfs/Gemini-Blog/image-Oct-24-2026-07-30-20-0452-AM.jpeg)](https://blog.blazingcdn.com/en-us/what-is-the-opus-codec-low-latency-audio-explained?hsLang=en-us) Learn ### [What Is the Opus Codec? Low-Latency Audio Explained](https://blog.blazingcdn.com/en-us/what-is-the-opus-codec-low-latency-audio-explained?hsLang=en-us) The Opus codec is a royalty-free audio codec, standardized by the IETF as RFC 6716 in 2012, that encodes speech and ... Oct 24, 2026, 9:31:57 AM [Read more](https://blog.blazingcdn.com/en-us/what-is-the-opus-codec-low-latency-audio-explained?hsLang=en-us) [![BlazingCDN](https://blog.blazingcdn.com/hubfs/Logo/blog-logo-w.png)](https://blog.blazingcdn.com?hsLang=en-us) [📘 Learn](https://blog.blazingcdn.com/cdn-learn?hsLang=en-us) [📊 Benchmarks](https://blog.blazingcdn.com/cdn-ratings-and-benchmarks?hsLang=en-us) [🎬 Video & Streaming](https://blog.blazingcdn.com/video-streaming-cdn?hsLang=en-us) [🏭 Industries](https://blog.blazingcdn.com/cdn-industry-insights?hsLang=en-us) [💰 Pricing & Costs](https://blog.blazingcdn.com/cdn-pricing-and-cdn-costs?hsLang=en-us) [⚡ Compare](https://blog.blazingcdn.com/cdn-comparison?hsLang=en-us) [🔒 Security](https://blog.blazingcdn.com/cdn-security?hsLang=en-us) [🛠️ Tools](https://blog.blazingcdn.com/cdn-tools?hsLang=en-us) [✍️ Publish with us](https://blog.blazingcdn.com/publish-with-us?hsLang=en-us) in f 𝕏 ✉ Copyright © BlazingCDN |. All rights reserved. ```json { "@context" : "https://schema.org", "@id" : "https://blazingcdn.com/#organization", "@type" : "Organization", "description" : "High-volume CDN for video, live streaming, OTT/IPTV, software, games, SaaS and large-file delivery.", "logo" : { "@id" : "https://blazingcdn.com/#logo", "@type" : "ImageObject", "height" : 560, "url" : "https://blazingcdn.com/wp-content/uploads/2024/08/logo-560-560.png", "width" : 560 }, "name" : "BlazingCDN", "sameAs" : [ "https://www.linkedin.com/company/blazingcdn", "https://x.com/BlazingCdn", "https://www.reddit.com/r/BlazingCDN/", "https://www.wikidata.org/wiki/Q141711074" ], "url" : "https://blazingcdn.com/" } ``` ```json { "@context" : "https://schema.org", "@id" : "https://blog.blazingcdn.com/#website", "@type" : "WebSite", "inLanguage" : "en-US", "name" : "BlazingCDN Blog", "publisher" : { "@id" : "https://blazingcdn.com/#organization" }, "url" : "https://blog.blazingcdn.com/" } ``` ```json { "@context" : "https://schema.org", "@id" : "https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained#article", "@type" : "BlogPosting", "author" : { "@id" : "https://blog.blazingcdn.com/en-us/author/blazingcdn#person", "@type" : "Person", "description" : "Vadym Borovskyi is Chief Product and Marketing Officer at BlazingCDN, where he has led product strategy and growth since 2021. He has more than 11 years of experience in IT product management and B2B marketing. On this blog he writes about CDN pricing and egress costs, video and software delivery, cache configuration, and how teams evaluate and switch CDN providers. He is based in Warsaw, Poland.", "image" : "https://143144902.fs1.hubspotusercontent-eu1.net/hubfs/143144902/vadym-borovskyi.jpg", "jobTitle" : "Chief Product and Marketing Officer", "knowsAbout" : [ "Content Delivery Networks (CDN)", "CDN pricing", "Video streaming delivery", "Software and game distribution", "Cache optimization", "Product Management", "B2B Marketing", "Digital Marketing", "Pay Per Click Marketing (PPC)", "Google Ads", "Search Engine Marketing (SEM)", "Marketing Strategy", "Artificial Intelligence (AI)", "Cloud Management", "SaaS Development", "IT Consulting", "Computer Networking", "Project Management", "Startups" ], "name" : "Vadym Borovskyi", "sameAs" : [ "https://www.linkedin.com/in/axelt/" ], "url" : "https://blog.blazingcdn.com/en-us/author/blazingcdn", "worksFor" : { "@id" : "https://blazingcdn.com/#organization" } }, "dateModified" : "2026-10-25T08:32:22Z", "datePublished" : "2026-10-25T08:32:22Z", "description" : "Understand VMAF, PSNR, and SSIM—key video quality metrics—and learn how PSNR vs SSIM compare for video assessment.", "headline" : "Video Quality Metrics: VMAF, PSNR and SSIM Explained", "image" : "https://143144902.fs1.hubspotusercontent-eu1.net/hubfs/143144902/Gemini-Blog/image-Oct-25-2026-08-30-18-1503-AM.jpeg", "inLanguage" : "en-us", "isPartOf" : { "@id" : "https://blog.blazingcdn.com/#website" }, "mainEntityOfPage" : { "@id" : "https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained", "@type" : "WebPage" }, "publisher" : { "@id" : "https://blazingcdn.com/#organization" }, "wordCount" : 2137 } ``` ```json { "@context" : "https://schema.org", "@id" : "https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained#breadcrumb", "@type" : "BreadcrumbList", "itemListElement" : [ { "@type" : "ListItem", "item" : "https://blog.blazingcdn.com/en-us", "name" : "Blog", "position" : 1 }, { "@type" : "ListItem", "item" : "https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained", "name" : "Video Quality Metrics: VMAF, PSNR and SSIM Explained", "position" : 2 } ] } ``` ```json { "@context" : "https://schema.org", "@id" : "https://blog.blazingcdn.com/en-us/video-quality-metrics-vmaf-psnr-and-ssim-explained#faq", "@type" : "FAQPage", "mainEntity" : [ { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "A VMAF score of 93 to 95 on the default model is generally treated as visually transparent for a 1080p top rung on a TV. Lower rungs commonly sit between 60 and 85. Interpret scores relative to the model used: the phone model rates the same encode higher, and the 4K model assumes a closer viewing distance." }, "name" : "What is a good VMAF score for streaming?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "SSIM usually tracks viewer perception better than PSNR because it compares local structure instead of raw pixel error. Both still miss temporal artifacts and correlate weaker with viewer ratings than VMAF. PSNR remains the codec research baseline, while SSIM is useful for catching blocking and structural damage during encoder tuning and regression tests." }, "name" : "Is PSNR or SSIM better for measuring video quality?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "About 6 VMAF points is the commonly used just-noticeable-difference threshold for 1080p content on a TV. Differences of 2 points or less are rarely visible to viewers. Many teams use the 6-point spacing to place bitrate ladder rungs, which keeps each step visibly different while avoiding redundant renditions that add encode and storage cost." }, "name" : "How many VMAF points is a noticeable difference?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "VMAF compares frames pixel by pixel, so the distorted video must match the reference resolution. Upscaling a 720p rendition to 1080p with a fixed bicubic filter mimics what the player does on a 1080p screen. Changing the scaler between runs can move VMAF by several points, so keep the filter constant across every comparison." }, "name" : "Why do I need to upscale before computing VMAF?" } ] } ```