PureJsImage · JPEG XL

JPEG XL JavaScript library comparison

Dated public-API measurements across PureJsImage, jSquash, jxl-oxide-wasm and wasm-vips. Compare practical workflows and their limits.

These tools use the current repository build. Repository functionality may be newer than the npm release. Static lossy encoding is qualified; lossy animation remains Experimental.

Quick answer

One-page campaign summary

The October 2026 lossy campaign is complete on the first-party JavaScript encoder. It improves compression against the previous main and meets the approved speed entry limits. Overall compression parity remains open: the final photo ladder and reserved holdout both miss the wasm-vips SSIMULACRA2 and Butteraugli 3-norm mean targets. Full-band coverage limits the conclusions. Further compression tuning is paused by the owner.

Source fingerprint 1224c8a82ba91d08e127b1ca5f225a8cc8bb3d2d104152fa228e5b328139d7be. These effort-7 measurements use pinned eight-bit RGB/RGBA photo inputs. They do not establish HDR, float or CMYK compression parity. Final report and Complete campaign data.

The bounded-fallback correctness repair freshly reproduces all 1,253 normal promotion streams without increasing managed memory. The nine-distance compression scores retain their earlier producer and are bridged by identical encoded bytes. Formal speed and final eleven-distance reports are fresh on the repaired source. The single effort-9 screen retains its original producer and timings.

Photographic development, final eleven-distance ladder

64 frozen photographs, deterministic native-pixel crops, distances 0.45 through 14. Negative BD-rate means fewer bytes at equal measured quality. Mean values include measured overlaps; the full-band column is reported separately. SSIMULACRA2 uses 60–90. Butteraugli max uses 0.5–3. The 3-norm integral uses measured overlap, and its full-band flag requires 0.5–3.

Final photo BD-rate against pinned peers
PeerMetricMean BD-rateMeasured imagesFull band
jsquashSSIMULACRA2-1.08%6447
jsquashButteraugli max+15.63%6441
jsquashButteraugli 3-norm+15.60%640
vipsSSIMULACRA2+1.18%6446
vipsButteraugli max+6.63%640
vipsButteraugli 3-norm+9.87%640

The final eleven-distance report adds 10 and 14 after promotion. It stays separate from the nine-distance promotion and holdout ladder. Photo report, coverage and reproduction.

Reserved holdout, aggregate only

63 photographs and 567 fresh own points. Only aggregate quality results were inspected, and no further tuning followed. The SSIMULACRA2 target is at most 0% against both peers. Butteraugli targets use wasm-vips 0.0.19, libjxl 0.12.0: at most +5% for 3-norm and +8% for max. jSquash 1.3.0 remains a reference column.

Aggregate holdout on the nine-distance promotion ladder
PeerMetricMean BD-rateFull band / measuredImages above +20% on measured overlap
jsquashSSIMULACRA2-1.37%10 / 632
jsquashButteraugli max+13.46%30 / 6311
jsquashButteraugli 3-norm+10.40%0 / 634
vipsSSIMULACRA2+0.39%9 / 633
vipsButteraugli max+7.55%0 / 637
vipsButteraugli 3-norm+7.47%0 / 634

Partial-range means and zero full-band admissions do not prove parity across the required band. Against unchanged main, the holdout means improve by 3.38% on SSIMULACRA2, 5.44% on 3-norm and 0.80% on max. Peer targets are final report targets; promotion used the separately approved main-baseline limits.

Fresh isolated speed and memory

Formal effort-7 speed presets are our distance 2, jSquash quality 80 and wasm-vips distance 2 on identical prepared inputs. Encode times exclude perceptual scoring. Preset timing comparisons do not establish equal quality; quality-matched compression is assessed separately through BD-rate.

Crop median: 2.916× jSquash and 4.167× wasm-vips, with all 192 jobs freshly measured. Original matched-set median: 2.544× jSquash on nine common inputs. All-ten original median: 6.440× wasm-vips. The 21 MP jSquash encode aborts and has no timing in either matched median.

The 13 MP original is 2.270× jSquash, within the 5× target and 10× entry bound. All ten own original encodes complete. Curve-sweep times are excluded from these formal ratios.

All 90 own watch settings and 20 scale settings fit the unchanged 137,514,476-byte encoder-managed ceiling. Watch maximum: 136,508,774 bytes. Managed storage, encoder RSS and scorer-tree memory are separate quantities. The repaired 12 MP fine-distance case preserves its stream while saving 42,806,175 managed bytes.

Minified JPEG XL bundles: 611,664 core bytes and 684,484 specialized bytes. Bundle size is report-only for this campaign.

Slow effort 9 and secondary images

Effort 9 retains the complete-alternative searches and is documented as slow. One untuned 16-photo screen passes both independent decoders on all 80 own streams. At the declared presets, distance 2 for us and quality 80 for jSquash, its fresh median time is 12.463× our effort 7 and 11.624× jSquash effort 9. Quality-matched BD-rate stays separate from these preset timings. Effort-9 BD-rate, speed and coverage.

56 non-photo images use a separate final eleven-distance report. They are not mixed into photo means. Constant-quality curves and unavailable intervals remain N/A. Secondary image report. The final report preserves failed ideas, reference failures, memory limits and remaining compression gaps.

Earlier matched-point checkpoints

These retained measurements describe earlier encoder sources and narrower studies. The final campaign report above is the current lossy baseline.

Overall lossy compression parity remains open. At matched Butteraugli 0.5, the qualified portrait curve uses 0.09% fewer bytes than pinned jSquash and 15.25% fewer than wasm-vips. The independent 12 MP photo also retains its passing fine comparison. That checkpoint left 3 confirmed portrait size gaps and 4 unresolved comparisons. Later measurements below confirm two additional Butteraugli-three gaps. Fine-photo public results and source bridge.

The latest public-package portrait curve uses an estimated 385,082 bytes at Butteraugli three, 6.93% more than jSquash and 4.77% more than wasm-vips. Its bracket width is 0.115960; all three brackets meet the unchanged 0.25-unit rule. Five confirmed portrait size gaps remain open. The current portrait SSIMULACRA2-70 comparison remains unresolved. Exact formatted source and package-byte bridges retain these results at source 170c6e66. Current coarse public comparison and validation. The preceding source-specific measurement remains historical.

Both complete Node files check all 57.6 million public/native RGBA samples within one code, with exact alpha, unchanged caller data and closed ownership. Real Chromium produces the exact same 389,618-byte Node file and checks all 28.8 million browser/native samples. The separate 12 MP photo keeps an identical 262,669-byte stream at distance 5.857; this is one-setting retention, without a new matched curve or size gain. Optional-stage low-memory fallback remains unqualified. Raw SSIMULACRA2 scores stay separate from the Butteraugli frontier. All four integrated package builds fit the adopted size ceilings.

The new large-transform public curve uses an estimated 675,638 bytes at matched portrait Butteraugli 2, 2.66% fewer than the preceding qualified public result. The bracket width is 0.198669, within the unchanged 0.25-unit limit. All 57.6 million public/native RGBA samples pass within one code, with exact alpha, unchanged caller data and closed ownership. The jSquash gap remains open. Public Butteraugli-two endpoints and peer curves.

Portrait Butteraugli 2; estimated PureJsImage file size versus frozen peers
PeerPeer bytesPureJsImage versus peer
jSquash553,69422.02% larger
wasm-vips724,4936.74% smaller

The protected portrait SSIMULACRA2-80 curve uses an estimated 770,030 bytes. Its bracket width is 0.095702, within the unchanged 0.25-unit limit. Both new endpoints pass all 57.6 million public/native samples, exact alpha, unchanged caller data and closed ownership. The current curve improves on the preceding result and both frozen peers. This does not close the remaining Butteraugli gaps. Public SSIM80 endpoints and separate reference curves.

Portrait SSIMULACRA2 80; current estimated file size versus separate references
ReferenceReference bytesCurrent PureJsImage versus reference
Preceding PureJsImage796,7313.35% smaller
jSquash820,0416.10% smaller
wasm-vips847,5599.15% smaller

The separate original 12 MP photo uses 6.78% fewer bytes at the same nominal encoder setting. Butteraugli changes from 1.9426 to 1.9463, and SSIMULACRA2 from 80.0296 to 80.0695. Its full public/native grid passes. This is one same-setting control; a matched independent-photo curve remains pending. The large-gradient control retains its existing quality and memory limits after independent native/Rust qualification. Independent public photo point and Large-gradient validation.

Qualified public files at matched Butteraugli 0.5; rounded byte estimates
InputPureJsImage bytesjSquash bytesVersus jSquashwasm-vips bytesVersus wasm-vips
2400 × 3000 portrait2,562,1272,564,3990.09% smaller3,023,09515.25% smaller
4000 × 3000 photo2,451,5852,484,9041.34% smaller2,989,28817.99% smaller

The portrait's matched SSIMULACRA2-90 estimate is 2,008,483 bytes, 2.53% below the preceding result.

Each metric keeps its own nondominated log-byte frontier, unchanged 0.25-unit bracket rule and no extrapolation. Frozen peers and all failed attempts remain in the evidence. Cached research files are excluded from public curves. Package ceilings and exports are preserved. This establishes no new speed or isolated RSS result. Fine checkpoint source SHA-256: eaf22d025e94c5c05e3427a8f5ec6a4d9cd637e0883823525851465d26c4b627.

Fine allocation requires variation in the existing source importance scores. Inputs with nearly uniform source scores keep the preceding fine-quality path. All six published photo settings remain admitted. Their original files, decoder grids and quality scores are reused through an exact admission-only source bridge. The corrected complete packages are measured separately and remain within the original size ceilings. No photo was reencoded and no photo metric was refreshed for this guard correction.

The qualified B-channel change reduces the portrait's matched Butteraugli-1 estimate by 1.00%, to 1,559,334 bytes. It remains 15.39% larger than jSquash and 8.40% larger than wasm-vips. The Butteraugli-2 gap against jSquash at this checkpoint is 25.36%. The later Butteraugli-three result above keeps its own source and files. Middle-photo public files and validation.

Both complete public files match all natural-order coefficients and non-order geometry from the scored research files. Their learned scan orders differ legally. An original harness assertion required identical orders; its failed receipts remain preserved. Separate saved-file validation confirms exact native PNG identity before either quality score is reused. Full checks, the unchanged generated photo regression and real Chromium coverage pass. Fine and coarse paths retain their preceding behavior. No independent-photo middle curve or new speed and RSS result is claimed. Current source SHA-256: 5edb762179ceb5bce7cc2513f536202dd417772eaab38d183f8abba46ec7ab6d.

The preceding middle/coarse curves reduce the portrait's matched Butteraugli-1 estimate by 9.41%, to 1,575,149 bytes. It remains 16.56% larger than jSquash and 9.50% larger than wasm-vips. The earlier original's Butteraugli-1 and SSIMULACRA2-90 comparisons also improve. The fine update above clears one of this checkpoint’s four confirmed gaps. Preceding middle/coarse curves, source pins and preservation controls.

Preceding middle/coarse curves on two original photos; rounded byte estimates
InputMetricTargetPureJsImage bytesChange from historical curveFrozen jSquash bytesFrozen wasm-vips bytes
2400 × 3000 portraitButteraugli11,575,1499.41% smaller1,351,3881,438,511
2400 × 3000 portraitButteraugli2694,0882.43% smaller553,694Unresolved
2400 × 3000 portraitSSIMULACRA280796,7312.66% smaller820,041847,559
4000 × 3000 photoButteraugli11,262,1771.09% smaller1,291,1651,396,748
4000 × 3000 photoSSIMULACRA2901,416,4520.59% smaller1,639,8971,686,911
4000 × 3000 photoButteraugli2482,3541.44% smaller521,032Unresolved
4000 × 3000 photoSSIMULACRA280500,0972.52% smaller564,888557,876

These scoped curves use 19 fresh complete public-package files, 38 complete native/Rust grids and 700,800,000 checked public samples, with exact alpha and unchanged caller data. Additional tundra and Earthrise controls retain the exact baseline. Selection may encode a second candidate and keeps the smaller complete file, with baseline recovery on optional working-storage failure. This establishes neither a speed nor a memory improvement. Both metrics keep their own frontier, the original 0.25-unit bracket rule and no extrapolation. The peer encoders were not rerun; unresolved peer estimates remain unresolved. Preceding scoped source SHA-256: e0976829f3fa41ac24ea0b338727d82065fda130f881a78154048f781c5a5b3a.

Overall lossy compression parity remains open. The historical unresized 2400 × 3000 portrait baseline uses 10.50–28.66% more bytes than jSquash at three separately matched Butteraugli targets, despite competitive SSIMULACRA2 results. The earlier 12 MP photo results below keep their separate scope. Broader photo results and remaining gaps and Verified portrait baseline.

The preceding median-luminance entropy change saves 0.13–0.56% on six fixed-setting controls across the portrait, earlier photo, tundra and Earthrise originals, with identical decoded pixels and both quality scores. Median-luminance entropy controls.

The spatial histogram change saves another 0.07–0.82% on five of those six controls. Earthrise keeps the preceding bytes. Every decoded pixel and both quality scores remain identical. This adds encoding work and does not close the portrait's matched-quality gaps. The historical tables below retain their earlier curve source; the latest scoped curves are shown separately. Spatial source SHA-256: 6e53c41aab93a239347ab47f46f5b4e06542948d9f2b29089d5dbecaa58ceee1. Spatial entropy preservation controls.

On the original 4000 × 3000 photo, PureJsImage uses 9–13% fewer bytes than jSquash and 8–16% fewer than wasm-vips at matched SSIMULACRA2 70, 80 and 90. It also uses fewer bytes at all four matched Butteraugli targets against jSquash and both resolved targets against wasm-vips; two wasm-vips targets remain unresolved under the unchanged rules. Parity applies to the resolved targets on this photo; other original-size photos, HDR, float, CMYK and animation need separate matched compression evidence. Original-size photo results and limits and Original checkpoint production controls.

Original 4000 × 3000 photo at matched SSIMULACRA2
TargetPureJsImage bytesjSquash byteswasm-vips bytesVersus jSquashVersus wasm-vips
70271,113302,681309,13910.43% smaller12.30% smaller
80513,039564,888557,8769.18% smaller8.04% smaller
901,424,8841,639,8971,686,91113.11% smaller15.53% smaller
Same original photo at independently matched Butteraugli targets; lower error is better
TargetPureJsImage bytesjSquash byteswasm-vips bytesVersus jSquashVersus wasm-vips
0.52,482,4432,484,9042,989,2880.10% smaller16.96% smaller
11,276,1151,291,1651,396,7481.17% smaller8.64% smaller
2489,411521,032Unresolved6.07% smallerUnresolved
3239,867249,307Unresolved3.79% smallerUnresolved

Sizes are log-byte interpolated estimates rounded for display. Both metrics use their own nondominated frontier, brackets no wider than 0.25 units, a fixed 24-attempt budget and no extrapolation. Separate matched metrics do not establish that one file dominates both metrics together. The wasm-vips budget is exhausted: its target-2 bracket is 0.27216 units wide and target 3 has no bracket. Public WASM failures remain in the raw study; native encoder sizes never fill these gaps.

At the preceding curve checkpoint, fresh complete public-package encodes reproduced all 14 selected candidate files exactly. The 74 selected native/Rust grids and 672 million public decoded samples were rehashed and counted as reused qualification. All 16 candidate curve points have fresh complete independent grids and 768 million checked public samples. Optional photo quantization and entropy search covers regular opaque standard-sRGB RGBA8 above four and through sixteen megapixels; working-storage failure retains the baseline and cancellation propagates. Both complete packages matched the qualified hashes and retained exports, growing by 3,531/3,734 bytes within explicitly updated 558,000/631,000-byte ceilings. Curve checkpoint source SHA-256: 7c17231f8d1c13b79eaf4794cdf69709160c9d962ed5501f5492da8a42387771. Raw points and failed attempts, Selected endpoint qualification, and Three additional original-size correctness controls without matched peer curves.

Both pinned 1024 × 768 photographs also use fewer bytes than both frozen WASM peers at matched SSIMULACRA2 70, 80 and 90. The remaining first-photo score-70 gap changes from 3.00% larger to 0.53% smaller than wasm-vips. All twelve capped-photo comparisons pass the same original bracket rule. Capped photo compression and boundaries and Preceding capped-photo measurements.

The separately rechecked opaque graphic produces 14,055 bytes at SSIMULACRA2 84.39 and 37,051 bytes with every original sample exact. These points dominate all six sampled peer files selected at targets 70, 80 and 90 on size and both quality metrics. This is a direct point comparison; the missing interpolation brackets and original comparison counts stay visible. The full exact-artwork search takes several minutes per encode, so this establishes compression results without a speed claim. Graphic point comparisons and Graphic production controls.

The preceding alpha palette search reduces the original transparency fixture from 48,819 to 9,038 bytes with identical decoded colors, exact alpha and unchanged quality scores. A higher-quality 9,155-byte point is smaller than the frozen 9,307-byte wasm-vips point and scores better on both SSIMULACRA2 and Butteraugli. This proves dominance of that measured point and keeps the frozen matched-pair counts unchanged. Alpha compression and boundaries and Alpha production measurements.

The fresh transparency comparison passes all six selected peer/target files on size and both aggregate quality metrics. Distance 6 uses 8,409 bytes against wasm-vips at 8,642, with SSIMULACRA2 88.34796 versus 88.10510 and Butteraugli 2.28504 versus 2.39888. Alpha stays exact. Scores use the worse result over black and white backgrounds; individual background results remain in the report. The selected jSquash files alter alpha, while the selected wasm-vips files preserve it. The photo, graphic and transparency studies cover all 24 original peer/target selections using matched capped-photo brackets or direct graphic/transparency files. Other original-size photos, HDR, CMYK and animation compression need separate evidence. Transparency point comparisons and Fresh transparency production controls.

The latest integer-lossless encoding closes all three remaining measured size gaps. Every one of the 45 independently exact comparable cells in the original Node matrix now meets or beats the frozen peer size. All four separate original-size photo and screenshot controls also meet or beat peer size. This is compression parity on the pinned integer-lossless corpus. Broader lossy, float, CMYK and animation compression remains open.

PureJsImage now decodes all 16 pinned lossless inputs exactly, including the seven that previously failed. It exposes native sample and exact JPEG workflows. The dated comparison below records toolkit sizes and runtime performance. The WASM alternatives are faster on several workloads. The historical quality table still has 11 missing adequate interpolation brackets; fresh direct graphic and transparency comparisons appear above.

Later unreleased compression changes reduce the same 4000 × 3000 lossless effort-1 photo from 15,389,193 to 7,452,825 bytes, versus the pinned 9,306,869-byte jSquash and 8,176,908-byte wasm-vips files. The preceding integrated DC model reduces 67 of the 91 original lossy files with identical decoded samples. The preceding photo filter changes 27 files, making them slightly larger at the same distance setting, with no SSIMULACRA2 regressions and ten Butteraugli regressions. Its separate cost studies pass the former 5% timing guard and authorized 10% peak RSS allowance. That historical photo source is 5abf439d5dbdb53ad73ce5a5c9c414497cc770055e288dcfdfa51c0e0213f138. The preceding capped-photo transform and entropy search closes the remaining measured capped-photo gap. The first photo is 0.55% smaller than wasm-vips at score 80 and 0.53% smaller at score 70. Every one of the 34 current photo files has complete native/Rust and public JavaScript pixel checks. The combined comparison has 13 adequately matched pairs and 11 unresolved pairs. The second photo's tighter 0.05-score search remains inconclusive; its separately recomputed 0.11437-score bracket passes the original 0.25 rule. That checkpoint raised package ceilings to 554,000/627,000 bytes; the current original-photo checkpoint raises them to 558,000/631,000. No speed or RSS parity is claimed. The speed tables below retain their earlier source. Current photo source SHA-256: ce3cc0a731d114c4a5f0400fec1fe27ef228ecb874010c28db8676e20b4ef149.

The earlier reversible-color checkpoint shrinks four more of the sixteen original files and leaves twelve byte-identical to the preceding checkpoint. None grows. Bounded reversible-color search reduces the map from 289,879 to 268,214 bytes, versus 268,611 for wasm-vips and 301,588 for jSquash. All four capped graphics below are now smaller than both frozen peers. All native samples, transparent RGB and alpha stay exact across 64 complete independent decoder grids. This checkpoint prioritizes compression first. Its independently audited 21-pair map study measures 37.20% more cold encoding time, 38.07% more warm time and 16.94% more cold peak RSS than the preceding PureJsImage checkpoint. The previous cost-guard failures remain recorded. Original working-memory budgets recover their former streams, and package ceilings stay unchanged. Every preceding effort-7 file and all 56 native color encodings remain byte-identical. Reversible-color checkpoint source SHA-256: 239bf2414446881c31cb4a5936c1fa56b55e8e11ed38dfbb3e4b3e5edb7689df. Reversible-color compression and remaining gaps, Reversible-color production controls, and Preceding spatial checkpoint.

The preceding effort-7 checkpoint compares denser prediction training while retaining the previous complete group. Twelve original files shrink, four remain byte-identical, and none grows. All original native samples, transparent RGB and alpha remain exact in both independent decoders. The map falls from 133,365 to 125,561 bytes, below both frozen peers. All effort-1 files and all 56 native color encodings remain byte-identical. Original working budgets recover their exact previous streams. This adds search work and makes no paired speed claim. Eighteen changed-path cases pass in Chromium, Firefox and WebKit. Denser training checkpoint source SHA-256: 37243714e4990afd617240afcdcd1a305a79f6d820420ec7ea308e057849fe06. Denser lossless prediction checkpoint and Preceding production controls.

Current effort-1 lossless graphics, 1024 × 1024, bytes; each encoder retains its original settings
InputFormer PureJsImageCurrent PureJsImagePinned jSquashPinned wasm-vips
im26-5032-diagnostic289,879268,214301,588268,611
im26-5034-diagnostic135,150107,746202,254158,783
im26-5052-diagnostic119,79292,849217,711179,817
im26-5334-diagnostic47,38347,383135,588105,926

The latest optional group search retains every completed ordinary and patch file, then tries smaller groups with bounded local reversible color selection and a refined patch-display tree. Six effort-1 files and four effort-7 files shrink; none grows. The two remaining capped photo files become 819,166 bytes at effort 1 and 641,956 bytes at effort 7. The original screenshot becomes 386,813 bytes, below pinned jSquash at 391,394 bytes. Every original native sample, hidden RGB and alpha remains exact in both independent decoders. Existing memory limits, minimum boundaries, native color hashes, public exports and package ceilings remain preserved. Thirty changed-path cases pass in Chromium, Firefox and WebKit. This adds encoding work; no speed or RSS parity is claimed. Current lossless codec source SHA-256: 792b04814811f17458fca81c3c21ff661412f3731e0f68405e0312ba1934041f. Lossless group search and remaining scope and Current lossless production controls.

The original Node report has 45 independently exact comparable lossless cells. Current files are at or below the pinned peer size in all 45. Three further cells have no exposed matching competitor API and remain excluded. Four additional original-size effort-7 controls retain their separate scope and all meet peer size. Overall compression parity remains open.

Compression changes and controls, current lossless results, and prior full quality replay.

Photo filtering and AC compression results, Independent photo checkpoint controls, and earlier DC-only results record the latest photo endpoints, original memory limits, package ceilings and browser checks. The frozen peer encoders were not rerun.

Six additional wasm-vips files resolved two previously missing comparisons on the second 1024 × 768 photo. The integrated DC comparison uses those frozen peer estimates; the table has 13 adequately matched pairs and 11 unresolved pairs. The sampler refines the same nondominated curve used for matching, with the unchanged 0.25-score bracket limit. That peer supplement preserves the original reports and adds no PureJsImage encodes or speed measurements. Sampling correction and scope, Peer measurements and matching calculations, and Earlier five-point supplement.

The integrated opaque palette change improves the original 1024 × 1024 graphic. Its qualified 14,617-byte distance-8 file scores 84.390 in SSIMULACRA2 and 4.820 in Butteraugli, versus 25,580 bytes / 80.153 / 7.669 for one pinned jSquash file and 30,921 bytes / 80.459 / 5.574 for one pinned wasm-vips file. Actual candidate files beat all 24 measured wasm-vips graphic files and 23 of 24 jSquash files in size and both quality metrics. The jSquash quality-101 exception remains. These complete-file comparisons add no tight matched target or competitor speed result. Independent decoder checks, the original 123-file matrix and production tests in Node, Chromium, Firefox and WebKit pass. The runtime tables below retain their earlier source. Palette results and limits.

Current effort-7 lossless files, bytes; effort numbers select each encoder's own policy
InputPureJsImagePinned jSquashPinned wasm-vips
4000 × 3000 photo6,183,4756,899,1886,648,964
1920 × 1080 screenshot386,813391,394422,389

At the preceding checkpoint, the effort-7 search reduced the original photo from 6,220,181 to 6,183,475 bytes and the screenshot from 401,990 to 400,951 bytes. At that checkpoint, the screenshot was 2.442% larger than pinned jSquash. Its two remaining main lossless gaps were 2.049% for im26-2018 at effort 1 against wasm-vips and 0.294% for im26-1416 at effort 7 against jSquash. Every original sample remains exact in both independent decoders. Preceding replay controls and Earlier effort-7 results and source identities.

Measured 2026-10-01 on one Linux workstation. PureJsImage base revision 590bbf08968f0f12c32bb4aa5c84627ba5dbeab4 with local codec changes. Codec source SHA-256: 088ba519315f3498cd388eaf0c7bef92c0a15c6702107e04219ed486760332ca. These results measure a local repository build.

Additional fixture checks now cover multi-group global Palette and Squeeze streams, all 27 VarDCT transforms, grouped progressive alpha, and ordinary ICC and structured SDR conversion through 16-bit integer precision. Relative custom white points and primaries also convert to sRGB. Ordinary IEEE float input preserves source samples, converts supported ICC or structured SDR color to straight sRGB16, and supports explicit PQ/HLG/linear HDR conversion. Ordinary lossless output preserves binary32 samples and matching retained ICC profiles. Native Modular float and CMYK frames now compose before display conversion, including selected animation frames. Separate libjxl, LittleCMS and FFmpeg references pass in Node and three browsers. Further independent checks cover 24 mixed-alpha and wide-integer layouts, custom float fields, RGB LUT8/LUT16 XYZ profiles, relative custom HDR, integer ICC preservation, Float32 animation writing and selected HDR VarDCT frames. HLG and PQ forward writing support associated alpha and nondefault luminance targets. HDR VarDCT reference frames compose in linear color before the output transfer curve. Explicit lossy Float32 output uses Modular mantissa rounding with exact alpha. Independent native groups decode cropped bands. Remaining gaps include transformed implicit-palette prefix layouts, nonrelative custom color, wider ICC/CMYK profile families, wider integer VarDCT display, float VarDCT writing and progressive float-alpha or animation output. The runtime tables retain the earlier source, inputs and comparator versions. The current compression results above use a later repository build and preserve their separate source identity. These additional capability checks cover cases beyond that shared comparison cohort. See the codec support checklist for the supported scope.

Correctness and API coverage

Chromium main set: 16 still inputs, with 13 lossy decode fixtures. Counts retain failures. Encode counts include multiple settings. “Incorrect” means the declared exact-sample or RGB8 tolerance contract failed, not a claim that the library is generally broken.

Verified / attempted public workflows; all other outcomes retained
LibraryOperationVerified / totalAPI absentUnsupportedIncorrectExecution failureUnavailable / not tested
PureJsImage repository builddecode-lossless16 / 1600000 / 0
PureJsImage repository builddecode-lossy13 / 1300000 / 0
PureJsImage repository buildencode45 / 4500000 / 0
PureJsImage repository buildroundtrip1 / 100000 / 0
@jsquash/jxl 1.3.0decode-lossless12 / 1630100 / 0
@jsquash/jxl 1.3.0decode-lossy4 / 1300900 / 0
@jsquash/jxl 1.3.0encode42 / 4530000 / 0
@jsquash/jxl 1.3.0roundtrip1 / 100000 / 0
jxl-oxide-wasm 0.12.6decode-lossless16 / 1600000 / 0
jxl-oxide-wasm 0.12.6decode-lossy13 / 1300000 / 0
jxl-oxide-wasm 0.12.6encode0 / 45450000 / 0
jxl-oxide-wasm 0.12.6roundtrip0 / 110000 / 0
wasm-vips 0.0.19decode-lossless16 / 1600000 / 0
wasm-vips 0.0.19decode-lossy13 / 1300000 / 0
wasm-vips 0.0.19encode45 / 4500000 / 0
wasm-vips 0.0.19roundtrip1 / 100000 / 0
Native, animation, metadata and access probes
LibraryFeatureStatusScope / evidence
purejsimageordinary wide integer preservationverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageordinary integer color with floating alphaverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageordinary Float32 lossless encodingverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageordinary integer ICC preservationverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimagefloating VarDCT and linear HDR reference blendsverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimagenative float samplesverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageexplicit display conversionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageHDR headroom 1/2/4verifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageextra channel samples and meaningverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageICC preservationverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageanimation encodeverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageanimation decodeverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageanimation compositionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageframe timingverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageframe accessverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageexact JPEG recompressionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageoriginal-byte reconstructionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageincremental input (known-length read-at source)verifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageregion reconstructionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimagereduced resolutionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageearly progressive outputverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageselective source readsverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
purejsimageExif/XMP preservationnot testednode/browser · unavailable
purejsimageorientation renderingnot testednode/browser · unavailable
purejsimagearbitrary animation blend modesnot testednode/browser · unavailable
purejsimagemultithread speedupnot testednode/browser · unavailable
purejsimagelossy native 16-bit encode/decodenot testednode/browser · unavailable
jsquashnative integer above 8-bitAPI not exposedpublic API inspection · survey.json
jsquashnative floatAPI not exposedpublic API inspection · survey.json
jsquashextra channelsAPI not exposedpublic API inspection · survey.json
jsquashanimation decodeAPI not exposedpublic API inspection · survey.json
jsquashframe timingAPI not exposedpublic API inspection · survey.json
jsquashframe accessAPI not exposedpublic API inspection · survey.json
jsquashanimation encodeAPI not exposedpublic API inspection · survey.json
jsquashICCAPI not exposedpublic API inspection · survey.json
jsquashmetadata accessAPI not exposedpublic API inspection · survey.json
jsquashincremental inputAPI not exposedpublic API inspection · survey.json
jsquashearly progressive outputAPI not exposedpublic API inspection · survey.json
jsquashreduced resolutionAPI not exposedpublic API inspection · survey.json
jsquashregion reconstructionAPI not exposedpublic API inspection · survey.json
jsquashselective source readsAPI not exposedpublic API inspection · survey.json
jsquashexact JPEG recompressionAPI not exposedpublic API inspection · survey.json
jsquashoriginal-byte reconstructionAPI not exposedpublic API inspection · survey.json
jsquashExif/XMP preservationnot testednode/browser · unavailable
jsquashorientation renderingnot testednode/browser · unavailable
jsquasharbitrary animation blend modesnot testednode/browser · unavailable
jsquashmultithread speedupnot testednode/browser · unavailable
jsquashlossy native 16-bit encode/decodenot testednode/browser · unavailable
oxideframe timingincorrect outputnode · benchmark/jpegxl/comparison/results/specialized-node.json
oxideanimation decodeverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
oxideanimation compositionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
oxideframe accessverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
oxideincremental input (sequential feed)verifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
oxideearly progressive outputverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
oxideregion reconstructionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
oxideICC accessverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
oxidestatic encodeAPI not exposedpublic API inspection · survey.json
oxideanimation encodeAPI not exposedpublic API inspection · survey.json
oxidenative raw integer planesAPI not exposedpublic API inspection · survey.json
oxidenative raw float planesAPI not exposedpublic API inspection · survey.json
oxideextra channelsAPI not exposedpublic API inspection · survey.json
oxidereduced resolutionAPI not exposedpublic API inspection · survey.json
oxideselective source readsAPI not exposedpublic API inspection · survey.json
oxideexact JPEG recompressionAPI not exposedpublic API inspection · survey.json
oxideoriginal-byte reconstructionAPI not exposedpublic API inspection · survey.json
oxideExif/XMP preservationnot testednode/browser · unavailable
oxideorientation renderingnot testednode/browser · unavailable
oxidearbitrary animation blend modesnot testednode/browser · unavailable
oxidemultithread speedupnot testednode/browser · unavailable
oxidelossy native 16-bit encode/decodenot testednode/browser · unavailable
vipsnative float samplesverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsnative-light HDRverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsanimation decodeverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsframe timingverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsanimation compositionverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsframe accessverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsICC accessverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsanimation encode with exact 3/5 ms delaysincorrect outputnode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsanimation encode at ordinary 30/50 ms delaysverifiednode · benchmark/jpegxl/comparison/results/specialized-node.json
vipsincremental inputAPI not exposedpublic API inspection · survey.json
vipsearly progressive outputAPI not exposedpublic API inspection · survey.json
vipsreduced-resolution JPEG XL reconstructionAPI not exposedpublic API inspection · survey.json
vipsselective source readsAPI not exposedpublic API inspection · survey.json
vipsexact JPEG recompressionAPI not exposedpublic API inspection · survey.json
vipsoriginal-byte reconstructionAPI not exposedpublic API inspection · survey.json
vipsExif/XMP preservationnot testednode/browser · unavailable
vipsorientation renderingnot testednode/browser · unavailable
vipsarbitrary animation blend modesnot testednode/browser · unavailable
vipsmultithread speedupnot testednode/browser · unavailable
vipslossy native 16-bit encode/decodenot testednode/browser · unavailable
vipsregion reconstructionnot testednode/browser · unavailable

All four initialized in Node and Chromium, with representative Firefox and WebKit checks. Lossy animation in PureJsImage remains Experimental. Exact 3/5 ms timing is a separate edge case: wasm-vips normalizes it to 100/100 ms; its 30/50 ms control passes. Oxide timing fields did not match the expected timeline; its frame access and composition passed.

Current quality and compressed size

13 of 24 PureJsImage/comparator/target comparisons have measured brackets no wider than 0.25 score points on both sides. The rest remain unresolved. The photo checkpoint freshly encoded 91 settings across four diagnostics at SSIMULACRA2 70/80/90 and retains Butteraugli coordinates, monotonicity evidence and missing intervals. Competitor curves retain their pinned versions and settings. Alpha is scored on black and white; comparator default alpha distortion remains visible. Exact graphic output does not substitute for a missing target bracket.

Current adequately matched size estimates; PureJsImage divided by comparator
InputSSIMULACRA2ComparatorSize ratio
im26-1030-diagnostic70jsquash0.988×
im26-1030-diagnostic80jsquash0.989×
im26-1030-diagnostic90jsquash0.911×
im26-1030-diagnostic70vips0.995×
im26-1030-diagnostic80vips0.994×
im26-1030-diagnostic90vips0.892×
im26-1416-diagnostic70jsquash0.940×
im26-1416-diagnostic80jsquash0.948×
im26-1416-diagnostic90jsquash0.760×
im26-1416-diagnostic70vips0.839×
im26-1416-diagnostic80vips0.850×
im26-1416-diagnostic90vips0.724×
alpha_triangles90jsquash0.743×

These capped diagnostics do not establish original-size or HDR compression parity. All measured points, brackets and unresolved pairs.

Actual source crop from the earlier measured photo
source · normalized source
Actual purejsimage crop from the earlier measured photo
purejsimage · measured SSIMULACRA2 80.037, setting 3.2469
Actual vips crop from the earlier measured photo
vips · measured SSIMULACRA2 79.983, setting 1.7586

Earlier comparison crops at identical 256 × 192 coordinates, shown without rescaling. They use actual nearby measured points rather than synthesized target-80 images and retain the earlier source. Source: lilith, PD-own, public domain dedication in imazen-26. Crop hashes and provenance.

Earlier paired lossless size and warm-time ratios
Runtime / comparatorScope / effortPairsMedian bytes ratioMedian warm-time ratio
chromium / jsquashcapped / 181.04711.641
chromium / jsquashcapped / 781.2425.586
chromium / jsquashoriginal / 150.9932.922
chromium / vipscapped / 181.15125.360
chromium / vipscapped / 781.1319.574
chromium / vipsoriginal / 151.0186.378
chromium / vipsspecialized / 130.7081.211
node / jsquashcapped / 181.04713.360
node / jsquashcapped / 781.2426.742
node / jsquashoriginal / 150.9933.141
node / vipscapped / 181.15129.190
node / vipscapped / 781.13111.503
node / vipsoriginal / 151.0188.387
node / vipsspecialized / 130.7081.130

These retain the earlier source used for the speed measurements. PureJsImage divided by comparator, verified pairs only. Complete originals, capped diagnostics and precision fixtures stay separate. Equal effort numbers do not mean equal work.

Warm encoding and startup

Representative capped photo, lossless effort 1, Node. Three warm operations after initialization, outputs materialized and independently checked. These are exact-output comparisons, not equal numeric lossy distances.

purejsimage713.0 ms [703.8, 716.9]
jsquash54.5 ms [53.2, 55.3]
vips27.8 ms [26.6, 28.2]

Oxide exposes no encoding API, so it has no encoding time.

Cold initialization plus first encode

Same Node photo workflow; median [minimum, maximum], three fresh processes
LibraryStatusCold millisecondsPeak process RSS, MiB
purejsimageverified773.4 [763.8, 774.9] (n=3)210.2 [177.2, 212.3] (n=3)
jsquashverified131.0 [124.0, 143.5] (n=3)179.7 [179.6, 179.7] (n=3)
oxideAPI not exposedunavailableunavailable
vipsverified200.0 [197.2, 203.7] (n=3)168.2 [167.5, 170.1] (n=3)

Cold reports separately include initialization and first operation, with three fresh observations for selected cells. Process/worker harness startup and input fetching are excluded. Warm operations reuse initialized modules. Small differences on this interactive host are not durable wins.

Oxide's decoder returns PNG, so its measured workflow includes mandatory PNG export. No raw-decode speed ranking equates that workflow with the other outputs.

Startup assets

Loaded encode/decode toolkit, required JS, WASM and worker assets counted once. Includes the common benchmark worker; these are not minimum decode-only application bundles. Brotli quality 11 is an offline estimate; actual requests used uncompressed no-store responses.

purejsimage124.3 KiB Brotli · 464.1 KiB raw
jsquash623.9 KiB Brotli · 2215.1 KiB raw
oxide422.7 KiB Brotli · 1674.1 KiB raw
vips2202.0 KiB Brotli · 7294.1 KiB raw

wasm-vips uses concurrency(1), but still starts its pthread runtime and requires cross-origin isolation. Actual repeated worker-script transfers remain in the raw report. Native libjxl is an oracle only and never enters JavaScript runtime comparisons.

Preview I/O and memory

PureJsImage's scaled region and oxide's full-frame PNG preview are different workflows; their latency is not ranked. The report retains first nonempty output, requested source bytes and transferred bytes. First-preview visual fidelity was not independently scored. Inspect actual stages in the local tool.

First nonempty preview, Chromium, median [minimum, maximum]. Different outputs, no speed ranking.
LibraryMillisecondsTransferred source bytesWorkflow
purejsimage59.9 [59.8, 60.6] (n=3)20544.0 [20544.0, 20544.0] (n=3)HTTP Range source, 32x32 region at scale 2; first nonempty pixel block; no canvas upload; geometry-only preview check
jsquashunavailableunavailablePublic incremental preview API not exposed
oxide61.9 [59.1, 66.9] (n=3)4096.0 [4096.0, 4096.0] (n=3)Sequential 4096-byte HTTP chunks; first render exported to PNG; full frame, no reduced-resolution API, no canvas upload
vipsunavailableunavailablePublic incremental preview API not exposed

Node peak RSS is isolated per subject process and includes the runtime, prepared source, output, WASM and worker infrastructure. It is not encoder-managed storage. WASM linear memory was not separately instrumented. Browser memory is unavailable across the tested isolation combinations. Canvas memory was not measured.

Methodology, uncertainty and downloads

Common inputs are pinned sRGB8 straight-alpha samples with orientation 1; high-depth samples are separate. Lossless validation includes color beneath zero alpha. PureJsImage's default lossy alpha must be exact; other wrappers' public lossy defaults may alter alpha, with that error retained. Native float probes and headroom 1/2/4 are not a perceptual HDR compression study.

Missing brackets, API boundaries, untested features and reference-limited comparisons remain visible. These inspected development sources are regression evidence, not an unseen holdout. The full comparison was rerun after the sample, ICC/HDR, animation and native crop updates. The dated library survey retains the pinned comparator versions.