Skip to content
Tool · AI video quality

How does this video look to a person?

Built by Raul Vecchione

The classic answer is a viewer panel: people watch, score each clip 1 to 5, and the average is the MOS, mean opinion score. It is slow and expensive. The fast answers are metrics. PSNR and SSIM compare pixels with the original; VMAF is a model trained on panel scores. All three need the original, and none says what is wrong.

This page asks a vision language model to do the panel's job, one frame at a time, beside the metrics. My pipeline, 3 clips, two encoder settings, one run.

Run 13 October 2026 · model gpt-5.5-2026-04-23 · 45 frames scored · 45 calls

What the run showed: the metric measures, the model describes

VMAF points lost, CRF 24 to 28
2.8 to 5.0VMAF points lost, CRF 24 to 28on every clip, as the bitrate fell
most the model's average moved
0.08most the model's average movedon the same step, out of a 1 to 5 scale
model score, source clips
3.5 vs 4.2model score, source clipsHoliday dance at 640×360 vs Waterfall at 1920×1080

Moving from CRF 24 to CRF 28 cut the bitrate on every clip, and VMAF fell every time. The model's average barely moved. It scored the untouched source about the same as the most compressed copy.

What the model did track was the picture itself: the low-resolution clip scored lowest at every setting, the Full HD clip highest. On small encoder steps it is not a replacement for a reference metric.

Its value is the written check. It names the artifact and where it sits, and it needs no original to compare against. Use the metric to decide and the model to explain. Measuring that before trusting it is the point.

The clips

Choose a clip:

Holiday dance

A green-screen holiday video: dancers composited onto a living room and a snowy roof. Low resolution, already heavily compressed. 640×360, 29.97 fps, 81.04 seconds, 2403 frames, source at 153 kbps.

Both encodes came out bigger than this source. It was already compressed hard, and CRF sets a quality target, not a file size.

VersionBitrateVMAF (0 to 100)PSNRSSIMModel MOS (1 to 5)
Source153 kbpsreferencereferencereference3.50
CRF 24250 kbps95.342.6 dB0.9913.52
CRF 28185 kbps92.639.6 dB0.9853.44
VMAF per frame, Holiday dance, CRF 24 and CRF 28VMAF for 1 frame in 13 of 2403. CRF 24 averages 95.3 (lowest 91.2); CRF 28 averages 92.6 (lowest 84.8). Markers show the 5 frames the model scored: frame 241, model 3.4; frame 721, model 4.0; frame 1202, model 2.8; frame 1683, model 3.6; frame 2163, model 3.4.8090100MOS 3.4MOS 4.0MOS 2.8MOS 3.6MOS 3.4frame (1 to 2403)
CRF 24CRF 28Dashed verticals: the frames the model scored, with its CRF 28 score.Plotted 1 frame in 13, of 2403.

The 5 frames the model scored

  • Holiday dance, frame 241, CRF 28 encode

    Frame 241 of 2403 · CRF 28

    Model
    3.4
    VMAF
    92.4
    PSNR
    38.8 dB

    The frame shows moderate overall visual quality with acceptable fidelity but clear limitations. Compression artifacts are mild to moderate: some blockiness and texture smearing are visible in the darker window areas, curtain folds, wooden floor, and around fine details such as the Christmas tree decorations. Slight ringing/haloing appears along high-contrast edges, especially around the person, curtains, and gift boxes. Blur/detail loss is noticeable due to the relatively low resolution and possible scaling; facial features, clothing texture, garland detail, and tree ornaments appear softened. A fast-moving object near the lower left is heavily motion-blurred, though this may be motion-related rather than compression-only. Color reproduction is generally bright and stable, with strong reds and greens, but saturation appears somewhat artificial and the composited background has a slightly flat, synthetic look. Edge preservation is fair: major silhouettes remain readable, but fine edges show aliasing and softness. Overall fidelity is adequate for casual viewing, but the frame lacks fine detail and exhibits mild compression and scaling artifacts.

    Same frame: source 3.2, CRF 24 3.4

    Source, model 3.2. The frame has moderate visual quality with noticeable limitations. Compression artifacts are present, especially as slight blockiness and mosquito noise around high-contrast edges such as the curtains, fireplace decorations, tree ornaments, and the subject's outline. Some ringing is visible along sharp graphic edges, while banding is minimal but may be faintly present in smoother wall and floor areas. Detail retention is mixed: the background elements remain reasonably recognizable, but fine textures are softened, and the subject's face, clothing details, and small decorations lack crispness, likely due to resolution limits and compression. Edges are generally preserved but appear somewhat aliased and artificially sharp in the background, with visible compositing edge softness around the person. Color reproduction is saturated and festive but not fully natural, with reds and greens appearing slightly boosted; skin tones are acceptable but somewhat flat. Overall fidelity is fair: the image is clear enough for viewing, but compression, low fine-detail resolution, and compositing-related edge artifacts reduce perceived quality.

    CRF 24, model 3.4, VMAF 95.5. The frame has fair overall visual quality with a clean but somewhat low-resolution appearance. Compression artifacts are mild: there is little obvious macroblock breakup, though some fine ringing/aliasing is visible along high-contrast edges such as the curtains, window frames, the subject outline, and Christmas tree decorations. Banding is minimal, with relatively smooth wall and floor gradients. Detail reproduction is moderate; the main subject is recognizable but facial detail, fabric texture, and small decorations appear soft, likely due to scaling or source resolution limits. A bright moving object near the lower left is heavily blurred and partially blown out, reducing local clarity. Color reproduction is vivid and generally stable, with strong reds and greens, though saturation is slightly heightened and some highlights near the snow/fireplace and blurred object are clipped. Edge preservation is acceptable but not crisp, with slight jaggedness and compositing-like edges around the subject and foreground elements. Overall fidelity is adequate for casual viewing but lacks fine detail and sharpness expected from high-quality video.

  • Holiday dance, frame 721, CRF 28 encode

    Frame 721 of 2403 · CRF 28

    Model
    4.0
    VMAF
    92.2
    PSNR
    38.5 dB

    The frame shows generally good visual quality with moderate sharpness and stable color reproduction. Compression artifacts are minimal, with no strong blockiness visible in flat wall or floor regions; slight ringing/aliasing can be seen around high-contrast edges such as the curtains, tree ornaments, and the subject outline. Fine details in the background, especially the snowy windows, tree decorations, and garland, appear somewhat softened, suggesting mild blur/detail loss or lower source resolution. Colors are vivid and well-saturated, particularly reds and greens, without obvious color bleeding or severe banding, though the overall image has a slightly artificial composited look. Edge preservation is mostly good on the main subject, but some thin lines and small decorations lose definition. Overall fidelity is good, with only minor softness and light compression/edge artifacts limiting the score.

    Same frame: source 3.6, CRF 24 3.6

    Source, model 3.6. The frame shows generally acceptable visual quality with moderate limitations. Compression artifacts are present but not severe: slight blockiness and mosquito noise can be seen around high-contrast edges such as the curtains, tree ornaments, fireplace decorations, and the subject's outline. Fine textures in the snowy window scenes, garland, and Christmas tree are somewhat smeared, suggesting compression and/or scaling loss. Blur is moderate; the main subject is recognizable but facial detail, clothing texture, and small decorative elements lack sharpness. Edge preservation is fair, with most major outlines intact, though some edges appear softened and slightly aliased, especially around the subject and vertical window/curtain boundaries. Color reproduction is vivid and stable, with strong reds and greens, but saturation appears somewhat artificial and the composite background has a flat, studio-like look. No major banding is visible, though darker areas near the bottom and in the outdoor window scenes show limited shadow detail. Overall fidelity is reasonable for casual viewing, but reduced fine detail, mild compression artifacts, and some edge softness keep it below high-quality presentation.

    CRF 24, model 3.6, VMAF 95.3. The frame shows generally acceptable visual quality with moderate limitations. Compression artifacts are mild to moderate, with some visible blockiness and texture breakup in detailed areas such as the tree, garlands, curtains, and window backgrounds. There is slight ringing/haloing around high-contrast edges, especially around the central subject and decorative elements. A small amount of banding or gradient compression is visible in the darker floor/shadow region near the bottom. Detail is somewhat softened, particularly in facial features, fabric texture, and background objects, suggesting limited resolution or compression-related detail loss. Color reproduction is vivid and stable, though reds and golds appear slightly oversaturated, and the overall image has a somewhat synthetic/composited look. Edge preservation is fair, but fine edges show aliasing and minor jaggedness, especially around the subject silhouette and ornaments. Overall fidelity is decent for a compressed video frame, but not high-definition quality due to softness, mild artifacts, and imperfect edge clarity.

  • Holiday dance, frame 1202, CRF 28 encode

    Frame 1202 of 2403 · CRF 28

    Model
    2.8
    VMAF
    93.9
    PSNR
    38.7 dB

    The frame shows moderate visual quality degradation. Compression artifacts are visible mainly as slight blockiness and mosquito/ringing artifacts around high-contrast edges such as the Santa figure, sleigh, snowflakes, and chimney. The dark sky background is mostly clean but shows minor noise/speckling and some unevenness; no severe banding is apparent. There is noticeable blur and detail loss on the central subject, especially in the face, beard, belt, and clothing texture, giving the figure a soft, low-resolution appearance. The snow texture retains contrast but appears repetitive and aliased, with jagged diagonal patterns and loss of natural fine detail. Color reproduction is generally vivid, with strong reds, yellows, and whites, but saturation appears somewhat artificial and the composited elements do not have consistent sharpness or tonal integration. Edge preservation is mixed: large graphic elements like the sleigh have fairly defined outlines, but many edges show jaggies, halos, and soft transitions. Overall fidelity is fair to low, with acceptable readability of objects but clear signs of compositing softness, aliasing, and compression-related edge artifacts.

    Same frame: source 3.4, CRF 24 3.2

    Source, model 3.4. The frame has moderate visual quality with limited obvious compression damage. There is slight ringing/mosquito noise around high-contrast edges such as the sleigh outline, Santa's red suit against the dark background, and the chimney, but strong blockiness is not apparent. Fine detail is limited: Santa's face and beard appear soft, the fur trim lacks texture, and the roof snow pattern looks somewhat smeared and repetitive, likely due to low resolution and scaling. Edges are generally preserved but show visible aliasing and stair-stepping, especially on the sleigh curves, roof ridges, and character outline. Color reproduction is vivid and stable, with saturated reds and yellows, though the composited elements have a slightly artificial look and inconsistent sharpness. Overall fidelity is fair: the image remains readable and stable, but softness, jagged edges, and minor edge artifacts reduce perceived quality.

    CRF 24, model 3.2, VMAF 96.8. The frame has moderate overall quality with clear visibility of the main elements, but several technical issues are present. Compression artifacts are relatively limited, though the dark sky shows slight noise/grain and minor block-like texture in flatter areas. High-contrast edges, especially around the sleigh, snow highlights, and character outline, show noticeable aliasing and some ringing-like harshness. The repetitive roof/snow pattern exhibits strong texture shimmer/moire-like detail and uneven preservation, which can appear visually busy and degraded. There is mild blur/detail loss on the central figure, particularly in the face and beard area, reducing fine detail definition. Color reproduction is vivid but somewhat synthetic, with saturated reds and stark whites; the whites retain detail in places but appear clipped or overly bright in some snow areas. Edge preservation is mixed: large graphic shapes remain sharp, but cutout boundaries and fine snow/star edges look jagged. Overall fidelity is fair, with acceptable clarity but visible compositing/aliasing artifacts and some loss of natural detail.

  • Holiday dance, frame 1683, CRF 28 encode

    Frame 1683 of 2403 · CRF 28

    Model
    3.6
    VMAF
    93.1
    PSNR
    41.2 dB

    The frame shows generally acceptable visual quality with moderate limitations from resolution and compression. Compression artifacts are mild to moderate: slight blockiness is visible in flatter areas such as the wall and floor, and there is some ringing/aliasing around high-contrast edges like the curtains, window frames, Christmas tree decorations, and the subject's outline. Fine detail is somewhat softened, especially in the face, hair, garland, tree ornaments, and presents, indicating blur/detail loss likely from compression and limited source resolution. Edge preservation is fair, with major object boundaries remaining clear, but thin structures and decorative elements appear jagged or smeared. Color reproduction is vivid and stable, with strong reds and greens, though saturation appears slightly exaggerated and the composited background has a somewhat artificial look. Overall fidelity is decent for a low-to-mid resolution frame, but fine texture reproduction and clean edge rendering are noticeably limited.

    Same frame: source 3.8, CRF 24 3.7

    Source, model 3.8. The frame has generally acceptable visual quality with no severe compression artifacts. Blockiness is minimal, though there is slight softness and possible mild ringing/haloing around high-contrast edges such as the curtains, the performer outline, and the tree decorations. Fine detail is somewhat limited due to the relatively low resolution and mild blur, especially in the face, hair, window background, and small ornaments. Color reproduction is vivid and stable, with strong reds and greens; saturation appears slightly exaggerated but not distorted, and there is no obvious banding in the walls or darker window areas. Edge preservation is fair: large object boundaries remain readable, but finer edges show some aliasing and compositing softness. Overall fidelity is good for a web/video frame, though detail clarity and edge precision are not high-end.

    CRF 24, model 3.7, VMAF 95.4. The frame shows generally acceptable visual quality with mild compression artifacts. There is some softness and detail loss on the central subject, especially in facial features, hair, clothing texture, and shoes, suggesting either scaling, focus softness, or compression smoothing. Fine details in the Christmas tree, garland, curtains, and outdoor window areas are partially preserved but appear slightly smeared in places. Compression artifacts are modest: minor ringing/haloing is visible around high-contrast edges such as the subject outline, window frames, curtains, and tree decorations, with slight blockiness in darker regions and flatter background areas. Banding is minimal, though smooth wall and shadow gradients show limited tonal subtlety. Color reproduction is vivid and stable, with strong reds and greens; saturation is high but not severely distorted. Edge preservation is fair, with major object boundaries remaining clear, though fine edges appear somewhat softened and compositing edges around the subject are not perfectly natural. Overall fidelity is moderate-to-good for a low-resolution frame, with the main limitations being softness, mild compression artifacts, and reduced fine-detail clarity.

  • Holiday dance, frame 2163, CRF 28 encode

    Frame 2163 of 2403 · CRF 28

    Model
    3.4
    VMAF
    90.3
    PSNR
    37.8 dB

    The frame has fair overall visual quality with a moderately clean image but noticeable limitations. Compression artifacts are mild: there is some visible softness and slight block-like texture in flatter areas such as the walls, floor, and curtains, with minor ringing/haloing around high-contrast edges like the figures, fireplace, stockings, and tree ornaments. Banding is not prominent, though gradients and shadowed regions appear somewhat simplified. Detail retention is moderate; the image looks downscaled or low-resolution, with facial features, beard texture, tree decorations, and background window scenery lacking fine detail. Motion or compositing blur is visible around the subjects, especially at limbs and costume edges. Color reproduction is generally vivid and stable, with strong reds and warm fireplace tones, though saturation appears somewhat exaggerated and the composite elements do not perfectly match in lighting or tonal depth. Edge preservation is acceptable but not crisp, with some jaggedness and cutout-like boundaries around the people and props. Overall fidelity is adequate for casual viewing, but limited by low spatial resolution, softness, mild compression artifacts, and visible compositing inconsistencies.

    Same frame: source 3.5, CRF 24 3.7

    Source, model 3.5. The frame shows generally acceptable visual quality with moderate limitations. Compression artifacts are present but not severe: there is some visible blockiness and mosquito noise around high-contrast edges such as the curtains, stockings, tree ornaments, and the Santa costume, with mild ringing/haloing around composited subject boundaries. Fine textures, especially the beard, fur trim, tree decorations, and background window detail, appear somewhat softened, indicating blur/detail loss likely from limited resolution and compression. Color reproduction is vivid and mostly stable, though reds are highly saturated and some bright areas in the fireplace and white fur show reduced tonal detail. Edge preservation is fair, but edges exhibit aliasing and cutout-like boundaries, particularly around the people and Christmas tree. Overall fidelity is decent for a low-to-mid resolution video frame, but softness, mild compression noise, and compositing edge artifacts prevent it from looking fully clean or natural.

    CRF 24, model 3.7, VMAF 93.9. The frame shows generally acceptable visual quality with moderate limitations. Compression artifacts are mild to moderate: slight blockiness and mosquito noise are visible around high-contrast edges such as the Santa costume, curtains, fireplace flames, stockings, and tree decorations, with some ringing/aliasing on fine decorative details. Blur/detail loss is noticeable, especially in facial features, the beard texture, background ornaments, and the snowy window area; fine textures appear softened, suggesting limited resolution or compression smoothing. Color reproduction is vivid and stable, with strong reds and warm fireplace tones, though saturation is somewhat exaggerated and highlights in white areas such as the beard, cuffs, and window snow lose subtle detail. Edge preservation is fair: major object boundaries remain clear, but composited edges around the people and decorations look slightly jagged or softened. Overall fidelity is decent for a low-to-mid resolution frame, but it lacks fine detail and has visible compression/processing artifacts that prevent it from appearing fully sharp or natural.

How it works

  1. 1. Encode

    FFmpeg re-encodes each source with libx264, preset medium, at CRF 24 and CRF 28.

  2. 2. Measure

    One FFmpeg pass per encode scores every frame against the source: VMAF (libvmaf, vmaf_v0.6.1), PSNR and SSIM.

  3. 3. Sample

    5 frames per clip, evenly spaced, the same frame numbers in every version.

  4. 4. Score

    A vision language model rates each frame alone, 1 to 5, and writes what it sees. It never sees the source or the metrics.

  5. 5. Publish

    Everything lands in one JSON file. This page reads that file; it never calls a model.

The prompt the model gets, unchanged since the first version
As a video quality expert, analyze this frame (#N of 5) for visual quality.

Please provide:
1. A Mean Opinion Score (MOS) between 1-5 (with one decimal place)
2. A detailed assessment of visual quality, covering:
   - Compression artifacts if present (blockiness, ringing, banding)
   - Blur/detail loss
   - Color reproduction issues
   - Edge preservation
   - Overall fidelity

Focus strictly on technical quality aspects visible in this specific frame, not content.
Format your response as a JSON object with fields: "mos" (number) and "assessment" (string).