You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
⚠️ BREAKING CHANGES
sqzer:Output has a new public field, lossless: the bytes hold exactly the samples the encoder was given. Code that builds an Output by hand sets it, and code that read Output::target to learn whether an encode was lossless reads the field instead (ADR-0012 D4)
metrics:Trial has a new public field, size: the bytes the trial encoded to. Search::run takes a closure that returns Scored, the output with its score and its size, in place of the pair of output and score. Code that builds a Trial or calls run adds the size
sqzer:Progress::Trial has a new field, size, the bytes the trial encoded to. A match that names every field adds it, or ..
sqzer:Target::Quality on a format set with Sqzer::format, or format in the npm package, whose encoder in this build only writes lossless is an error, where it encoded lossless and reported the quality: EncoderUnavailable with native-webp in available_in for WebP in the portable tier, as the README has said and the command line has done, and Unsupported for PNG, which no build writes lossy. Pass Target::Lossless instead, which is what the bytes were; a build with native-webp takes the quality (#56)
cli:--lossless and --preset lossless refuse -x png:optimize_alpha=true, which changes the colour under fully transparent pixels. Drop --lossless: PNG comes from a lossless-only encoder either way, so the file is the same. --json now reports such a run as "lossless": false
Features
codecs:-x png:colors=<2..=256> reduces a PNG to a palette, what rimage had as --quantization, through quantizr. -x png:dither=<0..=100> sets how much of the error is diffused, 100 by default. An image that already has that few colours is written exactly, at its own bit depth. The output is lossy: --json and the npm package say "lossless": false, {quality} in a --template renders lossy, and --lossless refuses the option (ADR-0012)
core:Encoder::exact says whether an encode at a lossless target keeps every sample under the given options. It defaults to true; a backend with an option that gives samples up overrides it, as oxipng does for png:colors and png:optimize_alpha
sqzer:Sqzer::lossy_options names the options set on the builder under which an encoder gives samples up, for a caller that wants to refuse a lossless target before encoding
cli: each entry of trials in a --json record carries size, the bytes that trial encoded to. -v prints it after the score, and the progress line shows it per trial
wasm: each Trial in trials, and what onTrial receives, carries size, the bytes that trial encoded to, so a page can plot quality against score and size as the search runs (#55)
wasm:options() lists the keys optimize and encode take, each with its default as the package spells it and its help, as codecs() lists the backend options: [{ key: "target", default: "70", help: "..." }, ...]. A key whose default depends on the image or the format says auto; one that does nothing when left out has none. The defaults are read from the library's builder, so a page can show every control at the package's default and drop its own copy (#57)
Bug Fixes
wasm: the npm package carries THIRD-PARTY-LICENSES.txt, the licence notice of every crate compiled into sqzer_bg.wasm, generated by cargo about from the wasm32 dependency tree as pack.sh builds the package, and the patent grant rav1e ships with its licence. The C library inside libdeflate-sys has its own MIT notice there, and avif-parse, under MPL-2.0, is listed with its repository. The package shipped its own two licence texts and nothing for the 170-odd crates in the module (#54)
cli: every release archive carries THIRD-PARTY-LICENSES.txt next to the binary: the notice of every crate compiled into it for that target, the C and C++ libraries of the native tier under the crates that vendor them, libwebp, libaom, libavif, libjxl and jpegli with the third-party code inside their trees, and the patent grants that travel with the AV1, WebP and JPEG XL licences. The archives shipped the project's own two licence texts and nothing else (#62)
sqzer: an explicit quality with no format, on a graphic the portable tier sends to lossless WebP, is reported as what it was: Output::target is Resolved::Lossless and Output::lossless is true, where the record said the quality asked for and lossless: false over the same lossless bytes. --json, {quality} in a --template and the npm package follow (#56)
Install sqzer-cli 0.4.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/sqzer-dev/sqzer/releases/download/v0.4.0/sqzer-cli-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
⚠️ BREAKING CHANGES
codecs:sqzer_codecs::png::PngEncoder, the plain png writer, is removed. sqzer_codecs::oxipng::OxipngEncoder writes PNG on every target; effort 0 is its fastest setting
Features
wasm: the browser build is sqzer on npm, built and staged from the release tag with provenance, live once approved (ADR-0011 D5): npm install sqzer. Its JavaScript API (ADR-0011 D3): optimize(bytes, options) for the whole pipeline in one call, decode and SqzerImage.encode to decode once and encode as often as a slider moves, decodeAny to hand what the package cannot read (SVG, HEIC where the browser has it) to the browser's canvas, fromPixels for pixels made elsewhere, and codecs() for what the build carries. Options are the command line's flags in camel case under the same rules, onTrial reports each trial of the target search, a result is the --json record in camel case with bytes added, and an error is a SqzerError whose kind is the sqzer_core::Error variant. The package is the portable tier minus svg, and its pixel limit defaults to 24 megapixels
sqzer: one feature per portable codec (png, jpeg, webp-lossless, avif, jxl-decode, heif, gif, tiff, bmp, tga, ico, qoi, pnm, exr, svg), forwarded to sqzer-codecs, for a build that leaves part of the tier out. portable is still all of them and still the default
codecs: PNG on wasm32 is written by oxipng, byte for byte what a desktop build writes from the same pixels, in place of the plain png writer, whose files were about 14 % larger. libdeflate is built there in its freestanding mode, so a wasm32 build of sqzer needs clang on PATH (ADR-0011 D1)
codecs: AVIF decodes on wasm32. The decoder is sqzer-rav1d on every target, in place of the archived re_rav1d: upstream rav1dmain with its safe Rust API and a stand-in for the libc names wasm32 lacks (ADR-0011 D2). A wasm32 build now reads AVIF input, takes a perceptual target for AVIF output and defaults to AVIF for photos, like every other build. --list-codecs and --json name the decoder rav1d
cli: errors and warnings style flags and values the way clap does: '--width' quoted, in yellow where it names the problem and in green where it names the fix. --color never, NO_COLOR and a pipe keep the quotes without the colour. The error and reason fields of --json carry the same text with backticks around each flag
Performance Improvements
workspace:wasm32-unknown-unknown builds of the workspace turn simd128 on through .cargo/config.toml. The target search runs about 1.7x faster there; a build of your own crate over sqzer sets the flag itself
Bug Fixes
workspace:wasm32-unknown-unknown builds of the workspace link with an 8 MiB stack through .cargo/config.toml. The linker's 1 MiB overflows in an unoptimised build of the decoders, which on wasm32 is a trap with no message. A build of your own crate over sqzer sets -C link-arg=-zstack-size itself
Install sqzer-cli 0.3.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/sqzer-dev/sqzer/releases/download/v0.3.0/sqzer-cli-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
⚠️ BREAKING CHANGES
core:Resize is the full request of ADR-0009 (size, fit, position, background, enlarge, filter) instead of max_width and max_height. Resize::inside(width, height) builds the old bounds, and Resize::fit returns an Option<Geometry> instead of an Option<(u32, u32)>; the output size is Geometry::output
sqzer:Sqzer::transform returns a Ready, and Sqzer::encode, encode_with and pick_format take one (ADR-0010). A caller that encoded a Decoded directly calls Sqzer::prepare then Prepared::resize, or Sqzer::transform, first
Features
cli: resize flags of ADR-0009: --width and --height with --fit inside|cover|contain|fill|outside in CSS object-fit terms, outside setting the shortest side with a square box (ADR-0010), --position for where cover crops and contain places the image, --background for the padding of contain (a translucent one needs a format with alpha, else exit 3), --scale 50%, --enlarge to allow scaling up, and --filter (lanczos3, mitchell, catmull-rom, bilinear, box, nearest). --max-width and --max-height behave as before
cli:--width 480,960,1600 writes one output per width and format from one decode, named photo-480w.avif and so on for srcset, each width searched on its own. Widths that come out the same size are refused before anything is encoded
cli: a dry run shows what a resize crops or pads, 48x32 -> 16x16 crop 32x32 at 0,0
cli: the rimage hint translates --resize specs into the new flags, --enlarge included unless the command passed --no-upscale
core: the resize geometry of ADR-0009 as pure functions: Resize::fit gives the crop of a cover fit, the resampled size and the canvas of a contain fit, for the dry run and the pipeline alike
sqzer: a type per stage, ADR-0010: Sqzer::prepare returns a Prepared once per input, Prepared::resize a Ready once per size, so several sizes come from one decode and one colour conversion. encode pads a contain fit with the background its encoder calls for, transparent where it takes alpha and white where it does not, and refuses a translucent one on an encoder without alpha with Error::Unsupported
Bug Fixes
cli:-f png,png, and a --template without {ext} or {format} under several -f formats, are argument errors instead of writing one file over another
Install sqzer-cli 0.2.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/sqzer-dev/sqzer/releases/download/v0.2.0/sqzer-cli-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
Features
codecs: HEIC input keeps its Exif and XMP items, the Exif orientation tag reset since the container's irot and imir are what rotate a HEIF
Bug Fixes
codecs: AVIF input applies the container's rotation and mirroring (irot, imir), keeps its ICC profile and its Exif and XMP items, and turns primaries other than sRGB from an nclx box or the AV1 sequence header (Display P3, BT.2020) into an ICC profile for the colour stage. Before, all of it was ignored: a rotated AVIF came out on its side and a P3 one with its colours shifted, even with --keep-icc or --keep-metadata
codecs: PNG input applies its EXIF orientation unless --no-auto-orient, and reads an eXIf chunk placed after the image data
Install sqzer-cli 0.1.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/sqzer-dev/sqzer/releases/download/v0.1.1/sqzer-cli-installer.sh | sh
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Release Notes
The first release. Everything below is new.
Features
cli: the sqzer binary, ADR-0003: one flat command, -f for one or several output formats, a perceptual target by default, -t, -q, --lossless, --preset (web, thumbnail, archive, lossless) and --fast to change it, -e for effort
cli: codec-specific options through -x codec:key=value only, checked against the backend before any file is touched. --list-codecs and --list-codecs -v print every backend in the build, its options and defaults, and which metadata it can carry
cli: output placement: a sibling file by default, -o file or directory, --suffix, --template with {stem} {ext} {width} {height} {format} {quality} {dir} {name}, --in-place with --backup, -r mirroring each input tree under -o. An output never overwrites its input unless --in-place, and one larger than its input is skipped unless --force
cli: inputs: an argument that exists is a literal path, one that does not and has glob characters is expanded in-process and case-insensitively. --files-from with -0, --include / --exclude under -r, - for stdin to stdout
cli:--max-width and --max-height: fit inside the bounds, keep the aspect ratio, never enlarge. The thumbnail preset fits inside 512 x 512
cli:--keep-icc and --keep-metadata. By default an ICC profile is converted to sRGB and dropped and EXIF and XMP are stripped; EXIF orientation is applied either way
cli:-j files in flight, bounded by an estimate of each file's peak memory, so a folder of large photos cannot exhaust the machine (ADR-0008). A file over the budget on its own runs alone with a warning
cli:--json writes one JSON Lines object per output and nothing else to stdout; -n decodes and plans without encoding; -v shows search trials, -vv resolved codec options. Exit codes 0 (done), 1 (partial batch), 2 (arguments), 3 (nothing could be done)
cli: coloured per-file lines, a summary line after a batch and a progress bar on a terminal, coloured by stage; --quiet, --progress, --color and NO_COLOR
cli: a format with no encoder in this build is an error naming the feature or the release archive that has one, never a fallback to another format. A rimage codec name as the first argument prints the equivalent sqzer line
sqzer:Sqzer, the library facade: decode, transform and encode, or run for the three in order, and encode_with to report progress to an observer. Output carries the bytes, the backend, its tier, the content class and the search report
sqzer: the perceptual target is the default: encoder quality is searched until the output scores SSIMULACRA2 70, in at most six encodes, starting from a seed table calibrated per backend
sqzer: a format picked by content when none is given: AVIF for photos, lossless WebP for graphics (PNG in a build without a WebP encoder)
sqzer: the colour stage over moxcms (ADR-0007), the resize stage over fast_image_resize (Lanczos3 in linear light, premultiplied alpha) and the metadata stage. Float input is converted to 16-bit sRGB, clipped at display white, no tone mapping
codecs: portable encoders, pure Rust apart from oxipng's vendored libdeflate (ADR-0002): JPEG over mozjpeg-rs, PNG over oxipng (the plain png writer on wasm32), lossless WebP over image-webp, AVIF over ravif
codecs: portable decoders: JPEG over zune-jpeg, PNG over png, WebP over image-webp, AVIF over re_rav1d (desktop only), JPEG XL over jxl-oxide, GIF over gif, TIFF over tiff, BMP, TGA, ICO, QOI and PNM over image's decoders, OpenEXR over exr and SVG through resvg
codecs: the native tier, ADR-0004, one feature per library: native-webp (lossy and lossless WebP over libwebp), native-jxl (JPEG XL over libjxl), native-avif (AVIF over libavif and libaom), native-jpegli (JPEG over jpegli) and native-heif (HEIC input through ImageIO on macOS, WIC on Windows and a runtime-loaded libheif on the other targets except musl, ADR-0005). A native encoder takes its format over from the portable one
codecs: the JPEG, WebP, JPEG XL, TIFF and HEIC decoders apply the image's orientation; PNG keeps EXIF but does not apply its orientation. Decoders keep an embedded ICC profile, except GIF, whose rare ICC extension is not read, and JPEG XL, which renders XYB and enumerated colour encodings to sRGB instead. JPEG, PNG, WebP and JPEG XL also keep EXIF and XMP, TIFF keeps XMP. The AVIF decoder reads none of ICC, orientation, EXIF or XMP, since avif-parse does not expose them. Encoders embed what they can carry and say so in EncoderCaps
metrics: SSIMULACRA2 over fast-ssim2 and Search, the target-quality bisection: a hard cap of six encodes, and a target the encoder cannot reach returns the best candidate with reached: false, never an error
metrics: seed tables for every lossy backend, portable and native, generated by tools/calibrate from CID22, GB82-SC and CLIC 2025. On a held-out split the search takes two to four encodes instead of four to six
core:Image with u8, u16 and f32 samples and ICC, EXIF and XMP on the struct; the Decoder, Encoder and Metric traits described by DecoderCaps and EncoderCaps; Registry, EncodeParams, Preset, Resize, Orientation, and Error with EncoderUnavailable and DecoderUnavailable
workspace: release binaries for six desktop targets with the native tier, built by cargo-dist (ADR-0006), plus shell and PowerShell installers and a Homebrew formula, brew install sqzer-dev/tap/sqzer. native on sqzer and sqzer-cli means every backend the target can build and run; a single native-* feature is strict
workspace: golden SSIMULACRA2 tests for every encoder, so a dependency bump that degrades output fails CI; CI on the six desktop targets and wasm32; cargo deny with a permissive licence allow-list
Install sqzer-cli 0.1.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/sqzer-dev/sqzer/releases/download/v0.1.0/sqzer-cli-installer.sh | sh
