# JPEG XL broader photo compression, October 8, 2026

## Quick answer

Overall lossy compression parity remains open. The latest qualified portrait
Butteraugli-0.5 curve is 0.09% smaller than pinned jSquash and 15.25% smaller
than pinned wasm-vips. The three earlier middle/coarse gaps remain. The latest public
coarse curve estimates 385,082 bytes at Butteraugli 3, leaving gaps of 6.93%
against jSquash and 4.77% against wasm-vips. Five confirmed portrait gaps
remain open. The new large-transform curve reduces the portrait's
Butteraugli-2 estimate by 2.66%, leaving a 22.02% gap against jSquash while
beating wasm-vips by 6.74%. The portrait's SSIMULACRA2-70 comparison still needs
current public files. The earlier 12 MP photo retains its passing fine result.
The protected portrait SSIMULACRA2-80 curve also improves by 3.35% and uses
6.10% fewer bytes than jSquash and 9.15% fewer than wasm-vips. The historical
tables below keep their original sources and unresolved cells.

The preceding median-luminance entropy change saves 0.13–0.56% on six fixed-setting controls across
four originals, with identical decoded pixels and both quality scores.
The [median-luminance controls](results/original-luma-context-controls.json)
keep their separate evidence.

The latest [spatial histogram change](results/original-spatial-context-controls.json)
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.
These changes add encoding work and do not close the portrait's matched-quality
gaps. The latest scoped curves below keep separate evidence from the
historical all-band baseline.

## Public large-transform update

Two complete public-package encodes at source SHA-256 `7fa4526a` qualify the
original portrait at Butteraugli 2. Distances 3.6905 and 4 produce 717,690 and
668,225 bytes, with Butteraugli 1.8320224285 and 2.0306913853. The bracket width
is 0.1986689568, within the unchanged 0.25-unit limit. The same nondominated
log-byte interpolation gives an estimated 675,638 bytes, 2.66% fewer than the
preceding qualified public result. Both frozen peer brackets also qualify.

| Input | Butteraugli | PureJsImage bytes | jSquash bytes | Versus jSquash | wasm-vips bytes | Versus wasm-vips |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| 2400 × 3000 portrait | 2 | 675,638 | 553,694 | 22.02% larger | 724,493 | 6.74% smaller |

The [public curve](results/large-transform-public-ba2.json) records complete
package and artifact identities, the actual endpoints and all source/input/tool
pins. All 57.6 million public/native RGBA samples agree within one 8-bit code;
alpha is exact, caller data is unchanged and managed storage closes. The two
raw SSIMULACRA2 scores are 78.59107954 and 77.17919958. No experimental cached
file enters this public comparison.

Two further public encodes qualify the protected portrait SSIMULACRA2-80 curve.
Distances 3.4158 and 3.3914 produce 769,571 and 774,430 bytes, with scores
79.99093737 and 80.08663964. Their bracket width is 0.09570227. The same
current-source log-byte interpolation estimates 770,030 bytes, 3.35% fewer
than the preceding qualified PureJsImage estimate of 796,731 bytes. It also
uses 6.10% fewer bytes than pinned jSquash and 9.15% fewer than wasm-vips.
All 57.6 million public/native RGBA samples pass the same one-code limit,
with exact alpha, unchanged caller data and closed ownership. The
[SSIM80 report](results/large-transform-public-ssim80.json) keeps all six actual
current points and separate historical references. Each metric uses its own
frontier. This result does not close the remaining Butteraugli gaps.

A separate full-search [12 MP photo point](results/large-transform-independent-point.json)
uses 467,236 bytes at distance 3.2613, 6.78% fewer than the historical
501,221-byte file at the same nominal setting. Butteraugli changes from
1.9426093102 to 1.9462729692, and SSIMULACRA2 from 80.02962323 to 80.06949255.
All 48 million public/native samples pass the same bounds. This single point
does not establish a matched-quality independent-photo curve.

The [large-gradient control](results/large-transform-gradient-control.json)
uses 43,823 bytes instead of 52,429. Complete native/Rust/public grids agree
within one code, alpha is exact, and mean public color error is 2.2590, below
the unchanged 2.35 limit. Managed peak storage remains within the original
64 MiB budget. Overall compression parity remains open.

## Public Butteraugli-3 update

Two complete public-package encodes qualify the original portrait at
Butteraugli 3. Distances 5.857 and 5.9471 produce 389,618 and 384,099 bytes,
with Butteraugli 2.9048168659 and 3.0207769871. Their bracket width is
0.1159601212. Both frozen peer brackets also pass the unchanged 0.25-unit
limit. The estimate uses the same nondominated log-byte interpolation without
extrapolation. Only these two actual public files enter the new curve.

| Input | Butteraugli | PureJsImage bytes | jSquash bytes | Versus jSquash | wasm-vips bytes | Versus wasm-vips |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| 2400 × 3000 portrait | 3 | 385,082 | 360,116 | 6.93% larger | 367,537 | 4.77% larger |

The [coarse public result](results/coherent-coarse-public.json) records exact
source and package bridges, actual artifact hashes and validation. The formatted
implementation reproduces all four qualified Node/browser package
bytes. Its measured codec and specialized packages fit the adopted 594,000 and
667,000-byte ceilings. The earlier prototype's package-budget failure remains
separate historical evidence. The source-trained coefficient model adds about
9 KB to the minified encoder package.

All 57.6 million Node public/native RGBA samples agree within one 8-bit code;
alpha is exact, caller bytes are unchanged and managed storage closes. Raw
SSIMULACRA2 scores are 63.53505915 and 63.15206802; no matched SSIM gain is
claimed. Real Chromium reproduces the 389,618-byte Node artifact exactly, and
fresh native pixels match the saved reference. All 28.8 million browser/native
samples pass the same bounds. Cancellation before source transforms and explicit
constructor memory rejection pass; optional-stage low-memory fallback remains
unqualified.

A separate 4000 × 3000 photo at distance 5.857 produces exactly the same
262,669-byte stream in the candidate and fresh baseline, with Butteraugli
2.9262928963 and SSIMULACRA2 69.53691064. Both files check all 48 million
public/native samples. This is one-setting retention, without a new independent
matched-quality curve or size gain.

The [preceding public comparison](results/original-portrait-ba3-public.json)
keeps its bc95029c/c262 source, distance-6.3/6.3645 files and 435,116-byte
estimate unchanged. Five portrait size gaps remain open: Butteraugli 1 against
both peers, Butteraugli 2 against jSquash, and Butteraugli 3 against both peers.
The current portrait SSIMULACRA2-70 comparison remains unresolved. No overall
lossy compression parity is claimed.

## Latest fine-photo update

The qualified portrait curve uses an estimated 2,562,127 bytes at matched
Butteraugli 0.5, 0.09% fewer than pinned jSquash and 15.25% fewer than pinned
wasm-vips. The independent 12 MP photo uses 2,451,585 bytes, 1.34% fewer than
jSquash and 17.99% fewer than wasm-vips. These are rounded log-byte interpolated
estimates from complete public files, with unchanged 0.25-unit brackets and no
extrapolation. Peer encoders were not rerun.

| Input | Butteraugli | PureJsImage bytes | jSquash bytes | wasm-vips bytes |
| --- | ---: | ---: | ---: | ---: |
| 2400 × 3000 portrait | 0.5 | 2,562,127 | 2,564,399 | 3,023,095 |
| 4000 × 3000 photo | 0.5 | 2,451,585 | 2,484,904 | 2,989,288 |

The [fine-photo controls](results/original-fine-production-controls.json) record
the qualified public files, source/package equivalence and protected SSIMULACRA2-90
comparison. The portrait's matched SSIMULACRA2-90 estimate is 2,008,483 bytes,
2.53% below the preceding result, 4.50% below jSquash and 10.40% below wasm-vips.
Four BA-0.5 endpoint files and two SSIM90 endpoint files have twelve complete
native/Rust grids and 211,200,000 checked public samples. Alpha is exact,
caller data is unchanged and encoder
and replay ownership closes. Public block releases and independent source-session
ownership evidence retain their separate scopes.

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.

At the fine checkpoint, three confirmed portrait size gaps remained: Butteraugli 1 against jSquash and
wasm-vips (15.39% and 8.40% larger), and Butteraugli 2 against jSquash (25.36%
larger). Four portrait comparisons were unresolved: jSquash SSIMULACRA2 70 and
Butteraugli 3, and wasm-vips Butteraugli 2 and 3. The later public Butteraugli-3
measurement above confirms two more gaps. No overall lossy parity or new
speed/RSS result is claimed. The separate cached equal-coverage diagnostic is
excluded from these public curves. Historical tables below keep their sources.

## Latest middle-photo update

The qualified B-channel change reduces the portrait's matched Butteraugli-1
estimate to 1,559,334 bytes, 1.00% below the preceding public result. It remains
15.39% larger than pinned jSquash and 8.40% larger than pinned wasm-vips.
The two complete public files bracket the target with a width of 0.18624,
within the unchanged 0.25-unit limit. The estimate uses log-byte interpolation
without extrapolation. Peer encoders were not rerun.

| Input | Butteraugli | PureJsImage bytes | jSquash bytes | wasm-vips bytes |
| --- | ---: | ---: | ---: | ---: |
| 2400 × 3000 portrait | 1 | 1,559,334 | 1,351,388 | 1,438,511 |

The [middle-photo controls](results/original-middle-channel-controls.json)
record both public files and their source and package identities. All 43.2 million
natural-order coefficients and non-order geometry match the scored research files.
The production encoder learns different legal coefficient scan orders. An original
harness assertion incorrectly required those orders to match; its failed receipts
remain preserved. Separate saved-file validation confirms exact native PNG identity
before reusing either quality score, without reencoding the photos.

The unchanged generated photo regression passes all public samples and both complete
native/Rust float grids. Full checks and real Chromium coverage pass. Fine distances
at most one and coarse distances at least two retain their preceding paths. This
does not establish an independent-photo middle curve, overall parity or new speed
and RSS results.

## Preceding middle/coarse curves

| Input | Metric | Target | Latest bytes | Versus historical curve | Frozen jSquash bytes | Frozen wasm-vips bytes |
| --- | --- | ---: | ---: | ---: | ---: | ---: |
| 2400 × 3000 portrait | Butteraugli | 1 | 1,575,149 | 9.41% smaller | 1,351,388 | 1,438,511 |
| 2400 × 3000 portrait | Butteraugli | 2 | 694,088 | 2.43% smaller | 553,694 | Unresolved |
| 2400 × 3000 portrait | SSIMULACRA2 | 80 | 796,731 | 2.66% smaller | 820,041 | 847,559 |
| 4000 × 3000 photo | Butteraugli | 1 | 1,262,177 | 1.09% smaller | 1,291,165 | 1,396,748 |
| 4000 × 3000 photo | SSIMULACRA2 | 90 | 1,416,452 | 0.59% smaller | 1,639,897 | 1,686,911 |
| 4000 × 3000 photo | Butteraugli | 2 | 482,354 | 1.44% smaller | 521,032 | Unresolved |
| 4000 × 3000 photo | SSIMULACRA2 | 80 | 500,097 | 2.52% smaller | 564,888 | 557,876 |

[Scoped production evidence](results/original-cone-floor-controls.json) records
19 fresh complete public-package files, 38 complete native/Rust grids and
700,800,000 public decoded samples. Every alpha sample and caller buffer
is preserved, and managed storage closes after each file. The formatted
production packages reproduce every measured compiled byte and keep the
original 558,000/631,000-byte ceilings.

Additional DCT16 choices apply to opaque standard-sRGB RGBA8 originals above
4,194,304 and through 16,777,216 pixels, at effort 7 and distances above one and
below four. Selection can encode an additional candidate and retains the smaller
complete file; optional storage failure retains the baseline. This adds work
and establishes no representative speed or RSS improvement.

The earlier photo's matched preservation gates pass. Previously observed
tundra and Earthrise controls keep the exact baseline bytes, pixels and both
scores. They do not supply a held-out or broader matched-curve claim.
Unmeasured bands, exhausted frozen peer budgets and every failed predecessor
remain separate. The peers were not rerun.

## Portrait baseline

This baseline uses production source `7c17231f`, public `@jsquash/jxl` 1.3.0
and `wasm-vips` 0.0.19 at effort 7. The input remains the exact original
opaque RGBA8 portrait, with no resize or crop. Each metric has its own
nondominated frontier, a maximum 0.25-unit bracket width and log-byte
interpolation without extrapolation. Byte counts below are rounded estimates.

| Metric | Target | PureJsImage bytes | jSquash bytes | Versus jSquash | wasm-vips bytes | Versus wasm-vips |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| SSIMULACRA2 | 70 | 527,895 | Unresolved | Unresolved | 536,955 | 1.69% smaller |
| SSIMULACRA2 | 80 | 818,469 | 820,041 | 0.19% smaller | 847,559 | 3.43% smaller |
| SSIMULACRA2 | 90 | 2,060,523 | 2,103,148 | 2.03% smaller | 2,241,666 | 8.08% smaller |
| Butteraugli | 0.5 | 2,833,640 | 2,564,399 | 10.50% larger | 3,023,095 | 6.27% smaller |
| Butteraugli | 1 | 1,738,709 | 1,351,388 | 28.66% larger | 1,438,511 | 20.87% larger |
| Butteraugli | 2 | 711,403 | 553,694 | 28.48% larger | Unresolved | Unresolved |
| Butteraugli | 3 | 438,361 | Unresolved | Unresolved | Unresolved | Unresolved |

All three participants exhaust their original 24-attempt budget.
jSquash's score-70 and Butteraugli-3 brackets remain too wide. Its measured
inversions stay in the evidence. Five wasm-vips bands are adequate; its
Butteraugli 2/3 brackets remain too wide. Those comparisons stay null.
An unresolved comparison does not imply a win or a loss.

The [physically verified baseline](results/original-matrix-portrait-baseline.json)
records complete native and Rust floating-point grids, exact alpha and every
PureJsImage public decoded sample. It rehashes the packages, encoded files,
scoring images, grids and tools, and freezes mutable source evidence. This
verification reuses already qualified measurements; it performs no new encode
or decode. It covers 72 files, 144 independent grids and 691,200,000 public
samples, with ten resolved and four unresolved comparisons. Four resolved
comparisons expose size gaps. The original comparison's 13 matched and 11 unresolved pairs stay
unchanged.

## Remaining work

The [predeclared plan](original-photo-matrix-plan.json) also includes the
unresized tundra and Earthrise originals. Their inputs are verified; matched
compression curves still need measurement. Encoder candidates must retain the
passing earlier photo results and close the confirmed portrait gaps before broader
parity is claimed. HDR, float, CMYK and animation require their own equivalent
public-API comparisons.

## Reproduction

Run these commands at the pinned baseline source, with the corpus and
development oracles prepared. Keep each heavy process tree inside the repository's
bounded benchmark runner.

```sh
node benchmark/jpegxl/comparison/prepare-original-photo-matrix.ts
node benchmark/jpegxl/comparison/refine-original-photo-matrix.ts .tmp/portrait-pure portrait-2400x3000 purejsimage
node benchmark/jpegxl/comparison/refine-original-photo-matrix.ts .tmp/portrait-jsquash portrait-2400x3000 jsquash
node benchmark/jpegxl/comparison/refine-original-photo-matrix.ts .tmp/portrait-vips portrait-2400x3000 vips
node benchmark/jpegxl/comparison/verify-original-photo-matrix.ts .tmp/portrait-verification.json portrait-2400x3000
```

Attempts persist across commands, including failures and interruptions.
Rerunning a command preserves the existing budget. Temporary images, grids,
packages and diagnostic scripts stay ignored.
