JPEG XL

JPEG XL (JXL) is the JPEG committee’s modern image format (ISO/IEC 18181). It does lossy and lossless coding well, handles HDR, wide gamut, high bit depth, alpha, layers and animation, decodes quickly, and is the only major format that can recompress an existing JPEG losslessly and give you the identical JPEG back. Browser support was the long-standing blocker; as of October 2026 that is changing quickly (see below).

Use it for archives and photo libraries (especially existing JPEG collections), for lossless images, for high-fidelity delivery, and anywhere you control the decoder. For lowest-bitrate web delivery, AVIF usually still produces smaller files.

Strengths and weaknesses

Strengths

• Lossless JPEG recompression. cjxl photo.jpg photo.jxl shrinks a typical JPEG by about 20 % and djxl photo.jxl photo.jpg restores the original file byte for byte.
• Strong lossless mode, usually well below PNG and lossless WebP on photographs, and it handles 16-bit properly.
• High-fidelity lossy. It is at its best at medium-to-high quality (distance 0.5–2), with graceful behaviour across a wide quality range.
• Fast decode, and encoding that scales with effort from near-JPEG speed to very slow.
• Progressive decoding, up to 32 bits per channel, HDR/wide gamut, many channels, tiled large images (over a billion pixels per side), layers and animation. Royalty-free.

Weaknesses

• Browser and application support is newer and patchier than for AVIF or WebP.
• At low bitrates (aggressive web compression) the best AVIF encodes are generally smaller and cleaner.
• Encoder tooling is essentially one implementation, libjxl. Its CLI and API still change between releases.

Support

Platform	Status as of 2026-10-08
Safari	Still images since Safari 17; animated JXL is not supported.
Chrome	Decoding is on by default from Chrome 155 (stable release scheduled 2026-10-06), using the Rust decoder jxl-rs. Chrome 145–154 had it behind #enable-jxl-image-format.
Firefox	Nightly only until Firefox 158, which enables it by default (stable scheduled 2026-10-13), also built on jxl-rs.
Edge, Opera, other Chromium	Follow their Chromium version, so expect support to spread over the following weeks.

Sources: Safari 17 notes, Chrome 155 beta and release notes, MDN’s JPEG XL data, Firefox 158 beta release notes and Mozilla’s intent to ship. Until Chrome 155 and Firefox 158 are widely installed, serve JXL through <picture> with an AVIF/WebP/JPEG fallback; image and audio support has the markup and a detection snippet.

Encode with cjxl

cjxl and djxl come from libjxl; most distributions package them.

cjxl input.png output.jxl -d 1

-d 1 and effort 7 are the defaults for PNG input. The two settings that matter:

• Distance (-d) is the lossy quality target in roughly just-noticeable-difference units. -d 1 is the default and is meant to be visually lossless at normal viewing; -d 0 is mathematically lossless; larger values are smaller and lossier (range 0–25, sensible range 0.5–3). -q is a JPEG-like alternative scale (-q 90 ≈ -d 1); use one or the other.
• Effort (-e) is encoder time versus density, 1–10, default 7. Higher effort mostly shrinks files slightly and makes the achieved quality more accurate; -e 9–10 are worth it for final lossless archives and rarely for lossy delivery.

Task	Command
Lossy, visually lossless	cjxl in.png out.jxl -d 1
Smaller web delivery	cjxl in.png out.jxl -d 2
Lossless	cjxl in.png out.jxl -d 0 -e 9
Lossless JPEG recompression	cjxl in.jpg out.jxl
Fresh lossy encode from a JPEG	cjxl in.jpg out.jxl -j 0 -d 1
Progressive loading	cjxl in.png out.jxl -d 1 -p
Photo-like grain on smooth images	cjxl in.png out.jxl -d 1 --photon_noise_iso=800
Animation from GIF/APNG	cjxl in.gif out.jxl
Decode to PNG	djxl in.jxl out.png
Restore original JPEG	djxl in.jxl out.jpg

JPEG input

A JPEG input is recompressed losslessly by default (--lossless_jpeg=1), keeping the reconstruction data so djxl can emit the original bytes. That is exactly what you want for archiving. If you instead want a new lossy JXL derived from the decoded pixels, pass -j 0 and a distance. Do not pass -d 0 -j 0 to a JPEG and expect the recompression: that is a lossless encode of the decoded pixels, and the file ends up several times larger than the JPEG.

Going the other way, djxl in.jxl out.jpg restores the original JPEG when the reconstruction data is present. djxl -j (--pixels_to_jpeg) forces a decode to pixels and a new, lossy JPEG encode, which is not a restore. Stripping metadata with -x strip=exif makes reconstruction impossible, and cjxl then requires --allow_jpeg_reconstruction=0.

Other options worth knowing

Option	Use
-m 0/1	Force VarDCT (0) or modular (1) coding; by default the encoder chooses.
-a DISTANCE	Alpha distance; default 0 (lossless alpha).
--keep_invisible=1	Keep RGB hidden under fully transparent pixels. It is on for lossless and off for lossy by default.
-p, --progressive_dc=0..2	More progressive/responsive decoding; --progressive_dc=1 adds an extra low-resolution pass (2 adds two).
--faster_decoding=0..4	Trade some density for lower decode cost on weak devices.
--photon_noise_iso=N	Synthesise film-like noise instead of encoding real grain (100 is a little, 3200 a lot).
-x color_space=…	Tell cjxl what colour space an untagged raw input is, or override one (sRGB, DisplayP3, Rec2100PQ, …).
--num_threads	Worker threads; default is the machine’s.
--container=1	Force the ISO container; usually only needed for metadata boxes.

Raw-pixel inputs (PPM, PFM and similar) carry no colour information; supply it with -x color_space=… or the output will be interpreted as sRGB. Animation inputs (GIF, APNG) are written as JXL animations with their frame timing; GIF input defaults to lossless (-d 0).

Compare JXL with the alternatives

• Against JPEG: clearly smaller at the same quality, and recompressing an existing JPEG saves about 20 % with no loss at all. jpegli narrows the gap for new lossy JPEGs.
• Against AVIF: AVIF is generally smaller at web-delivery quality; JXL encodes faster, decodes cheaply, holds up better at high quality and is much better in lossless.
• Against WebP: JXL wins in lossy at medium-to-high quality and in lossless, has no 8-bit/4:2:0 limits, and WebP has the support advantage.

For your own comparisons, SSIMULACRA2 and Butteraugli are the useful metrics, but libjxl is itself tuned with Butteraugli, so look at the images at 1:1 as well.

Pitfalls

• -d and -q are mutually exclusive.
• -e 10 is far slower than -e 9 for little gain; on camera-resolution images it is only worth trying for lossless archives.
• Check that the people who need to open the files have a JXL-capable viewer before converting a whole library; browser support is new and desktop and mobile apps vary.
• Re-encoding lossy JXL or lossy JPEG output with a lossy encoder compounds loss; keep sources.

Other implementations

• libjxl-tiny is a stripped-down encoder for photographic images meant to guide hardware implementations.
• Hydrium is a fast, very low-memory streaming encoder in C.
• zune-jpegxl is a simple Rust encoder with lossless 8- and 16-bit output.
• jxl-oxide and jxl-rs are spec-conforming Rust decoders; jxl-rs is what Chrome and Firefox use.

Related pages

• JPEG is what JXL most directly replaces, and what it can losslessly absorb.
• AVIF is the AV1-based alternative with better low-bitrate compression.
• WebP is the older widely supported web format.
• PNG is what lossless JXL most often replaces.

Full option reference

The table below is the full cjxl parameter reference for libjxl 0.12.0. Reading option tables explains the Default column.

cjxl 0.12.0 complete parameter reference

INPUT to —dec-hints

Parameter / aliases	Type or accepted values	Default	Purpose and constraints
INPUT	path	required	Source image: JXL, PNG/APNG, GIF, JPEG, PNM/PFM/PAM or PGX.
OUTPUT	path	required unless —disable_output	Output .jxl path; an existing file is overwritten.
--distance, -d	float 0–25	0 for JPEG/GIF input; 1 otherwise	Quality target in JND units; 0 is lossless, 1 is visually lossless. Excludes —quality.
--quality, -q	float 0–100	unset; input-dependent distance applies	Libjpeg-like quality mapped to distance; 100 is lossless, 90 is visually lossless.
--effort, -e	integer 1–10; 11 with expert opt-in	7	Encoder effort; higher is slower and usually smaller. 11 needs —allow_expert_options.
--version, -V	action	not requested	Print the encoder and library version, then exit.
--quiet	flag	false	Suppress progress and size messages.
--verbose, -v	repeatable flag	0 repetitions	More output; repeat up to four times (-v -v -v -v) to also list advanced options in —help.
--alpha_distance, -a	float 0–25	0	Distance for the alpha channel, 0–25; 0 keeps alpha lossless.
--progressive, -p	flag	false	Shortcut for progressive_ac, progressive_dc=1, center-first group order, no patches, and Squeeze.
--group_order	0 or 1	automatic: normally 0; progressive selects 1	Codestream group order: 0 scanline, 1 center-first. Center-first suits progressive loading.
--container	0 or 1	automatic when metadata/reconstruction requires it	1 forces the ISOBMFF container; 0 uses it only when metadata or JPEG reconstruction data needs it.
--compress_boxes	0 or 1	1	Brotli-compress metadata boxes when transcoding JPEG; 0 stores them uncompressed.
--brotli_effort	integer 0–11	9	Brotli effort for metadata boxes, 0–11; higher is smaller but slower.
--modular, -m	0 or 1	encoder chooses	1 forces Modular mode (lossless, graphics); 0 forces VarDCT (photos).
--lossless_jpeg, -j	0 or 1	1 for JPEG input; disabled for other inputs	1 transcodes JPEG input losslessly; set 0 to decode to pixels, required for lossy -d or -q on JPEG input.
--num_threads	integer -1 or nonnegative	-1: machine default	Worker threads; 0 disables multithreading and -1 uses all cores.
--photon_noise_iso	nonnegative float	0	Add film-like noise at this ISO (100 light, 3200 heavy); the output differs from the input.
--intensity_target	positive float, nits	0: choose from color encoding	Peak display brightness in nits for HDR content; an explicit 0 is rejected, so omit it for automatic.
--dec-hints, -x	KEY=VALUE, repeatable	none	Set or strip metadata, e.g. -x color_space=sRGB, -x icc_pathname=FILE, -x strip=exif; repeatable.

—allow_jpeg_reconstruction to —jpeg_reconstruction_cfl

Parameter / aliases	Type or accepted values	Default	Purpose and constraints
--allow_jpeg_reconstruction	0 or 1	1	Store data to restore the original JPEG bytes; 0 gives a smaller file that cannot restore them.
--codestream_level	-1, 5 or 10	-1: encoder chooses	Limit the codestream to level 5 (smaller limits, broadest decoder support) or 10; -1 chooses automatically.
--buffering	integer -1–3	-1: encoder chooses	Input buffering: 0 whole image (best compression), 1–2 streamed; use —output_mode instead of 3.
--faster_decoding	integer 0–4	0	Trade size or quality for faster decoding, 0 (off) to 4.
--premultiply	-1, 0 or 1	-1: preserve input association	Force alpha to straight (0) or premultiplied (1); -1 keeps the input’s.
--keep_invisible	0 or 1	1 for lossless; 0 for lossy	1 keeps RGB values under fully transparent pixels; 0 lets the encoder alter them for size.
--center_x	-1 or valid x coordinate	-1: image midpoint	Horizontal center for center-first group order; needs —group_order=1 or —progressive.
--center_y	-1 or valid y coordinate	-1: image midpoint	Vertical center for center-first group order; needs —group_order=1 or —progressive.
--progressive_ac	flag	false	Split AC coefficients into several passes so the image refines while loading.
--qprogressive_ac	flag	false	Progressive AC passes using shifted quantization instead of coefficient splitting.
--progressive_dc	integer -1–2	-1: encoder chooses	Extra low-resolution DC passes: 0 off, 1 adds a 64x64 pass, 2 adds 512x512 and 64x64.
--resampling	-1, 1, 2, 4 or 8	-1: input/quality-dependent	Downsample color channels by 1, 2, 4 or 8; -1 does so only at very low quality.
--ec_resampling	-1, 1, 2, 4 or 8	-1: encoder chooses	Downsampling factor for extra channels such as alpha; same values as —resampling.
--already_downsampled	flag	false	Input is already downsampled; skip encoder downsampling but tell the decoder to upsample.
--upsampling_mode	-1, 0 or 1	-1: nonseparable	Decoder upsampling: -1 smooth, 0 nearest neighbor (pixel art), 1 pixel dots at 4x and 8x.
--epf	integer -1–3	-1: encoder chooses	Decoder edge-preserving filter strength, 0 (off) to 3; -1 lets the encoder choose.
--gaborish	0 or 1	encoder chooses	Enable (1) or disable (0) the gaborish smoothing filter; unset lets the encoder choose.
--override_bitdepth	unsigned integer	0: input depth	Bit depth to write; 0 uses the input’s depth.
--noise	0 or 1	encoder chooses	Enable or disable the encoder’s adaptive noise synthesis; prefer —photon_noise_iso.
--jpeg_reconstruction_cfl	0 or 1	encoder chooses	Enable or disable chroma-from-luma when transcoding JPEG; unset lets the encoder choose.

—num_reps to —responsive

Parameter / aliases	Type or accepted values	Default	Purpose and constraints
--num_reps	positive integer	1	Encode this many times, for benchmarking.
--streaming_input	flag	false	Read PPM or PGM input incrementally to reduce memory use.
--streaming_output	flag	false	Write output through a callback as it is produced instead of buffering the whole file.
--output_mode	integer -1–2	-1: encoder chooses	Output layout: 0 buffered, 1 streamed with seeking, 2 out-of-order jxlp boxes without seeking.
--disable_output	flag	false	Encode without writing a file, for timing.
--dots	0 or 1	encoder chooses	Enable or disable dot detection and coding; unset lets the encoder choose.
--patches	0 or 1	encoder chooses	Enable or disable repeated-patch detection; unset lets the encoder choose.
--frame_indexing	0/1 index string	empty: no explicit index selection	Bit string marking which frames go in the frame index box, e.g. 1011.
--allow_expert_options	flag	false	Permit —effort=11, which is much slower for slightly denser lossless output.
--disable_perceptual_optimizations	flag	false	Turn off the encoder’s perceptual heuristics; listed only at help verbosity 4.
--iterations, -I	float -1–100	-1: encoder chooses	Percent of pixels used to learn modular MA trees; 0 disables the trees.
--modular_colorspace, -C	integer -1–41	-1: encoder chooses	Modular color transform: 0 none, 1–41 fixed reversible transform (6 is YCoCg).
--modular_group_size, -g	integer -1–3	-1: encoder chooses	Modular group size: 0 128, 1 256, 2 512, 3 1024 pixels per side.
--modular_predictor, -P	integer -1–15	-1: normally 14 below effort 10, 15 at effort 10	Modular predictor 0–15 (5 gradient, 6 weighted, 14 mixes 5 and 6, 15 mixes all).
--modular_nb_prev_channels, -E	integer -1–11	-1: encoder chooses	Maximum previous channels used as MA tree properties, 0–11.
--modular_palette_colors	integer -1 or nonnegative	-1: encoder chooses	Use a palette when the image has at most this many colors; 0 skips the palette search.
--modular_lossy_palette	flag	false	Use lossy delta palette; combine with —modular_palette_colors=0. Ignored in progressive mode.
--pre-compact, -X	float -1–100	-1: encoder chooses	Use a global channel palette when sample values are below this percent of the range.
--post-compact, -Y	float -1–100	-1: encoder chooses	Use per-group channel palettes when sample values are below this percent of the range.
--responsive, -R	-1, 0 or 1	-1: normally 0 lossless, 1 lossy; progressive selects 1	Apply the Squeeze transform for resolution-progressive decoding; default 0 lossless, 1 lossy.

—help and —

Parameter / aliases	Type or accepted values	Default	Purpose and constraints
--help, -h	action	not requested	Print help; add -v for advanced options (up to four times).
--	separator	not used	End of options; later arguments are file paths even if they start with a dash.