Image formats cheatsheet: JPEG, PNG, WebP, AVIF, JXL

Note: This article has been largely written by AI. Read this excellent AVIF vs JXL comparison by Jake Archibald instead.

TL;DR

• Best compression: AVIF and JPEG XL on par
• Best lossless compression: JPEG XL, then WebP
• JPEG XL can losslessly recompress existing JPGs by ~20%
• AVIF encoding is very slow in comparison
• Browser support for JPEG XL is evolving
• PNG, JPEG are outperformed by other formats in every way. WebP ranks behind AVIF&JXL in almost all categories

Lossless compression

• JXL (strong): usually the smallest lossless files, thanks to modern prediction and context modeling. Handles photos, graphics and high bit depth well. Typically ~30–50% smaller than PNG and also smaller than lossless WebP.
• WebP (good): clearly smaller than PNG, often 20–30%, usually a bit larger than JXL. Limited to 8-bit.
• AVIF (weak): built on AV1, which is designed for lossy video coding. The lossless mode works, but files are often as large as or larger than PNG or lossless WebP, and encoding is slow. It also needs an identity color matrix (4:4:4) to stay truly lossless, which many encoders and decoders handle poorly.

For archival, screenshots or graphics, AVIF is rarely worth choosing. JXL is the best fit when client support allows; lossless WebP is the broadly compatible option. Results vary by content, so measure on your own corpus.

Lossy compression

• AVIF (smallest at low bitrates): usually wins at aggressive compression, ~40–55% smaller than JPEG. Artifacts show up as smoothing and lost fine texture or grain. Encoding is by far the slowest.
• JXL (best at medium to high quality): on par with AVIF overall and often better when you want near-transparent quality. Preserves texture well; can show ringing at very low quality. Supports progressive decoding.
• WebP (good, weakest of the three): ~25–35% smaller than JPEG. Artifacts are blocking and chroma smearing, because lossy WebP is 8-bit and 4:2:0 only. Cheap to encode and universally supported.
• JPEG (baseline): the reference. Worst compression, but fast and universal. Optimized encoders like mozjpeg claw back some of the gap.

Compare formats at matched perceptual quality (for example SSIMULACRA2 or Butteraugli), not at the same quality slider value, since sliders mean different things per encoder.

Quality slider equivalents

If your tool only has a quality slider, use this table to find the setting that gives a similar visual result in each format. Pick a row for the quality you want, then read across to the slider value for your format. These are rough starting points, not guarantees.

• SSIMULACRA2: a perceptual score, 0–100, higher is better. It is the common scale that lets the formats be compared.
• JPEG, WebP, AVIF q: the usual 0–100 quality slider, higher is better.
• JXL -d: distance, lower is better, 1.0 ≈ visually lossless. The -q slider maps to it as d = 0.1 × (100 − q); the equivalent -q is in parentheses.

Example: for a "High" result, use JPEG or WebP at 90, AVIF at 70–80, or JXL at -d 1. The last column is the format that usually gives the smallest file at that level, ignoring client support.

Target	SSIMULACRA2	JPEG q	WebP q	AVIF q	JXL -d (q)	Smallest file
Low (thumbnails)	~50	60–70	60–70	40–50	3–4 (60–70)	AVIF
Medium (web default)	~65–70	75–80	75–80	55–65	2 (80)	AVIF, JXL close
High	~80	90	90	70–80	1 (90)	JXL, AVIF close
Visually lossless	~90	95+	95+	85+	0.3–0.5 (95)	JXL

Choosing between AVIF and JXL by quality target:

• AVIF below ~q60: smallest files; use AVIF.
• AVIF above ~q75–80: smoothing erases texture and encoding gets expensive; prefer JXL if clients support it, otherwise use AVIF with -y 444 and 10-bit.
• In between: close call, decide on client support.

Calibrate on your own images: encode a sample at several settings, score with ssimulacra2, and note each encoder's value at your target score.

LPIC's 2026 benchmark figures

These are figures reported by LPIC's 2026 comparison; treat them as results from that source, not universal format guarantees.

Its 3,200-image benchmark covered photos, UI screenshots, and illustrations. For a separate test, the article used the Kodak set of 24 photos at 768×512 and targeted SSIM 0.95.

Encoder	Encoded size	Encode time	Decode time
mozjpeg	100%	1.0×	1.0×
WebP	72–78%	3–5×	1.2×
AVIF (libaom, speed 4)	50–60%	50–100×	1.5×
AVIF (libavif, speed 8)	55–65%	8–15×	1.5×
JPEG XL (libjxl, effort 7)	55–65%	5–10×	1.1×
JPEG XL (libjxl, effort 3)	60–70%	1.5–3×	1.1×

Image feature properties

Property	JPEG	PNG	WebP	AVIF	JXL
Max bit depth	8	1–16	8	8 / 10 / 12	up to 32 (float)
HDR (PQ/HLG)	No	No (16-bit only)	No	Yes	Yes
Wide gamut / ICC	ICC	ICC	ICC, limited	Yes (CICP/ICC)	Yes, full
Alpha	No	Yes	Yes	Yes	Yes
Lossless	Rare	Yes	Yes	Yes (weak)	Yes (strong)
Animation	No	APNG	Yes	Yes	Yes
Progressive decoding	Yes (basic)	Interlaced (Adam7)	No	No	Yes (true progressive)
Max dimensions	65535×65535	~2 billion px per side	16383×16383	Very large (tiles)	~1 billion px per side
Chroma subsampling	4:2:0 / 4:4:4	None (RGB)	4:2:0 only (lossy)	4:2:0 / 4:2:2 / 4:4:4	Full control, XYB
Layers / extra channels	No	No	No	Limited	Yes
Depth maps / gain maps	Via extensions	No	No	Partial (gain maps emerging)	Yes

Numbers are rough; check the spec or encoder for exact limits.

When to use what

• JPEG: only where AVIF, JXL and WebP are not an option
• PNG: only where JXL and WebP are not an option
• WebP: pragmatic web default; good savings, cheap to encode, wide support. Also the broadly compatible lossless option.
• AVIF: best web delivery for photos at low bitrates, HDR. Pre-encode at build time because encoding is slow. Can look smooth and lose fine detail. Avoid for lossless.
• JXL: archival and master files, lossless (including lossless JPEG recompression), high-quality photos, progressive decoding. Use it where you control the decoder; add a fallback on the web.

Typical web setup: serve AVIF/JXL, fall back to WebP, then to JPEG or PNG, using <picture> or Accept header negotiation.