The image format landscape changes more slowly than the hype cycle suggests, and faster than most production teams notice. JPEG turned thirty in 2022 and still carries roughly seventy percent of all images on the public web. PNG, introduced in 1996, remains the default for screenshots, line art, and any image that benefits from lossless compression. WebP, launched in 2010, took a decade to reach majority browser support. The lesson across these histories is that adoption is bottlenecked by ecosystem inertia, not by encoder quality.

Three formats and one cross-cutting capability now drive the next adoption cycle: AVIF, which has already won the public web; JPEG XL, which dominates archival and professional pipelines; HDR-capable still imaging, which is moving from flagship phones into consumer expectations; and content provenance through C2PA, which is changing what an image file actually is. This article maps each in technical detail and offers concrete delivery guidance.

The Constraint That Shapes Image Formats

Image format design is constrained by a small set of physical and economic realities that have not changed since Lena was first scanned.

Human visual perception is non-linear and color-sensitive. Lossy compression succeeds by discarding information the eye does not register. The science of perceptual coding has improved, but the budget remains finite.

Decoder cost is paid once per byte, encoder cost is paid once per file. This asymmetry is why aggressive encoders ship and aggressive decoders do not. Every browser tab decodes thousands of images. Every encoder runs in a content pipeline that can afford CPU time.

Bandwidth is the primary cost on mobile. A 30 percent file size reduction across the public web translates directly to billions of dollars per year in carrier and CDN spend. This is why every major CDN now runs automatic format negotiation.

Patent and licensing risk shapes adoption. HEIC delivered better quality than WebP in 2017 and never reached the open web because of MPEG-LA licensing. AVIF reached the same quality, royalty-free, and adoption followed within three years.

"Compression research has been excellent for two decades. Compression deployment has been excellent for one. Closing that gap is mostly a question of patents, browsers, and tooling, not algorithms." Jyrki Alakuijala, principal author of JPEG XL and Brotli

What the Image Format Stack Looks Now

The current production stack is well understood and broadly stable. The following table captures the trade-offs as of 2026.

FormatBest UseCompression vs JPEGBrowser SupportNotes
JPEGUniversal fallback, photosbaseline100 percent8-bit only, no alpha
PNGScreenshots, UI, line artlarger lossless100 percentLossless, alpha, predictable
GIFLegacy animationmuch larger100 percentUse APNG or WebP instead
WebPWeb photos and graphics25 to 35 percent smaller98 percentLossy and lossless modes
AVIFWeb photos, HDR30 to 50 percent smaller95 percentSlow encode, fast decode
JPEG XLArchival, pro photo20 to 60 percent smallerSafari, Firefox flagLossless JPEG transcode
HEICiOS capturesimilar to AVIFSafari, native iOSPatent encumbered
SVGVector, iconsresolution independent100 percentXML, not raster
Two facts deserve emphasis. First, the formats with the strongest compression numbers (AVIF, JPEG XL) are also the formats with the most decoder complexity. Second, the gap between the best lossy formats and JPEG is now large enough that the bandwidth saving justifies the encoder cost in almost every production scenario.

AVIF: The Web's New Default

AVIF derives from the AV1 video codec and reuses its intra-frame coding tools. The format reached majority browser support in 2022 and is now the default output of every major image CDN. The compression advantage on photographic content is real, the decoder is fast on modern hardware, and the licensing story is clean (AOMedia patent pool with royalty-free terms).

A representative AVIF encode using libavif:

# Photographic content, web tier
avifenc --speed 6 --qcolor 50 --qalpha 60 \
        --yuv 420 --jobs all input.png hero.avif

# Pro tier with full chroma
avifenc --speed 4 --min 0 --max 30 \
        --yuv 444 --depth 10 input.png hero.avif

# HDR encode with PQ transfer
avifenc --speed 4 --depth 10 --yuv 420 \
        --cicp 9/16/9 hdr-input.exr hdr-output.avif

The CICP triplet (color primaries, transfer characteristics, matrix coefficients) is the part most teams get wrong. For SDR sRGB content the correct triplet is 1/13/6. For HDR10 the triplet is 9/16/9. Encoders that omit CICP produce files that display washed out on color-managed browsers, which is a frequent diagnostic for "AVIF looks worse than JPEG" complaints in production.

The HTML delivery pattern remains the picture element with three sources, ordered most-modern to most-compatible:

<picture>
  <source type="image/avif"
          srcset="hero-800.avif 800w, hero-1600.avif 1600w"
          sizes="(min-width: 768px) 1600px, 100vw" />
  <source type="image/webp"
          srcset="hero-800.webp 800w, hero-1600.webp 1600w"
          sizes="(min-width: 768px) 1600px, 100vw" />
  <img src="hero-1600.jpg" alt="Hero photograph"
       width="1600" height="900" loading="lazy"
       decoding="async" />
</picture>

This pattern degrades cleanly in every browser, including text-mode and accessibility tools. It is the right baseline even when CDN-side negotiation handles the same logic in HTTP headers, because it works without JavaScript and without an active CDN.

JPEG XL: The Archival Endgame

JPEG XL (ISO/IEC 18181) is the format that finally makes archival image storage rational. Three properties separate it from every prior format.

Lossless JPEG transcoding. A JXL encoder can take any existing JPEG and produce a smaller file (about 20 percent smaller on average) that is bit-exactly reversible to the original JPEG. No quality loss, no artifact stacking. This is the only format with this property.

Single codec for lossy and lossless. AVIF requires distinct encode paths. JPEG XL uses one variable transform pipeline. Lossless JXL files of UI screenshots routinely beat PNG by 20 to 35 percent.

Wide gamut and high bit depth as defaults. JXL handles 16-bit-per-channel HDR, alpha, animation, and ICC profiles in a single container. Multi-resolution decoding is built in, so a single file serves thumbnail and full-resolution requests from prefixes of the byte stream.

The deployment story is uneven. Safari 17, Firefox behind a flag, and most professional image tools (Adobe via plugin, DarkTable, Krita, ImageMagick, GIMP) ship JPEG XL support. Chrome removed support in 2023, citing insufficient ecosystem demand, then reopened the conversation under publisher pressure. As of 2026, Chrome support remains the missing piece for general web deployment.

"JPEG XL is not a marginal improvement. It is the first image format that simultaneously beats JPEG on size, beats PNG on size, and preserves bit-exact compatibility with the trillion JPEGs already in the wild. Refusing to ship that is a strategic error." Cloudinary engineering blog, 2023 review of image format options

For archive and pro-photo workflows, JPEG XL is the right format today. For public web delivery, the answer remains AVIF until Chrome ships JXL.

HDR Imaging Becomes Default

HDR (high dynamic range) capture is now standard on every flagship phone and most mid-tier devices. Apple's iPhone capture pipeline writes HDR HEIC by default, with metadata that drives gain map reconstruction in the photo viewer. Google's Ultra HDR format, introduced in Android 14, embeds a JPEG base image with a parallel gain map that HDR-aware viewers blend on display.

The implications for image formats are concrete:

  1. Capture pipelines now produce 10-bit-per-channel data, with PQ or HLG transfer functions and Rec. 2020 primaries.
  2. SDR rendering must come from explicit tone-mapping, not from clipping the high range.
  3. Browsers, OS compositors, and CDNs need consistent metadata to display HDR correctly.

The right encoder defaults for HDR AVIF or JXL include explicit CICP, embedded ICC profiles where the workflow demands them, and HDR-aware encoders (libavif 1.0+, libjxl 0.10+). Tools like ImageMagick now expose these knobs:

# Convert HDR EXR to AVIF with PQ transfer
magick input.exr \
  -depth 10 \
  -define avif:cicp=9/16/9 \
  -quality 90 \
  output.avif

# Inspect AVIF color metadata
exiftool -ColorSpaceTags -ICC_Profile output.avif
"HDR is not a brightness slider. It is a color-managed pipeline from sensor to display, and every step that loses metadata silently degrades the image. The format is the easy part, the discipline is the hard part." Dr. Lottie Westfall, color science engineer, Display Week 2024 keynote

For engineers building image pipelines, the practical move in 2026 is to assume HDR input from camera devices, preserve metadata across processing, and let the browser or display handle final tone mapping. The era of pre-baking SDR is ending.

Neural Image Codecs

Neural compression, in which a learned model replaces some or all of the encoder and decoder, has been a research staple since 2017. The benchmarks are striking: top neural codecs beat AVIF by 10 to 25 percent on perceptual metrics at equivalent file sizes. Production deployment lags, for reasons that are technical rather than political.

A neural decoder is a neural network. To display the image, every reader must execute that network. This creates four practical problems:

  1. Hardware variance. Different GPUs and NPUs produce subtly different outputs from the same model, which means a "lossless" neural codec is not lossless across devices.
  2. Decoder distribution. A new format must ship new code to every browser, OS, and image library. Neural decoders are large and version-sensitive.
  3. Model drift. Updating the model breaks every existing file. The format must lock in a specific model version, which forfeits the central advantage of learned compression.
  4. Energy. Neural decode draws meaningfully more power than DCT-style decode, which matters for mobile and laptop workloads.

The likely deployment path is narrow, hardware-specific use first. Camera-internal compression is the obvious wedge: Sony, Apple, and Samsung can ship a neural codec inside the device and convert to standard formats on export. From there, on-device generative upsamplers (already shipping in iPhone Photos and Pixel) blur the line between codec and post-processor.

For the public web, expect neural codecs to remain experimental through the late 2020s, with deployment first as quality enhancement layers on top of AVIF or JPEG XL rather than as standalone formats.

Provenance: C2PA and the Signed Image

The Coalition for Content Provenance and Authenticity (C2PA) defines a manifest format embedded inside JPEG, AVIF, JPEG XL, and other media containers. The manifest records:

  • The capture device or generating application.
  • An ordered list of edits and the tools that performed them.
  • Cryptographic signatures binding the manifest to a signing identity.

Sony, Leica, Nikon, and Canon all ship C2PA-compliant cameras. iPhone capture under iOS 18 writes C2PA-signed HEIC. Adobe Photoshop, Lightroom, and Premiere preserve C2PA across edits. The result is a property new to image formats: the file proves where it came from, not because of metadata that anyone can edit, but because of a chain of signatures that cannot be forged without breaking public-key cryptography.

The implications for image formats are structural. Every modern container now reserves space for JUMBF boxes (JPEG Universal Metadata Box Format), the carrier for C2PA manifests. Future format design assumes signed manifests are present, not optional.

"Provenance is a property of the bytes, not a property of the platform. A signed image survives copy, paste, and recompression. That is the only way provenance scales to the modern web." Andy Parsons, Content Authenticity Initiative, in a 2024 IETF presentation

Cross-domain pairings of provenance and content extend beyond images. Communities documenting biology and natural science, including the wildlife coverage at strangeanimals.info, are early adopters because contested provenance affects educational trust.

A Practical Delivery Stack for 2026

The right defaults for an image pipeline today are simple and survive at least the next three years.

Source storage. TIFF or JPEG XL for photographs. PNG or JPEG XL lossless for screenshots and line art. Preserve EXIF and ICC profiles. Hash and version every original.

CDN tier. Automatic format negotiation that emits AVIF for browsers that accept it, WebP for those that do not, and JPEG as a guaranteed fallback. Cloudflare Polish, Akamai Image Manager, Vercel image optimization, and self-hosted imgproxy or libvips pipelines all handle this.

HTML. The picture element with three sources. Width and height attributes always present, to avoid layout shift. Lazy loading on offscreen images.

Metadata. Preserve C2PA manifests through pipelines that handle camera output. Strip personally identifying EXIF (GPS, camera serial) from public-facing images unless the user opted in.

Monitoring. Track format distribution, average bytes-per-image, Largest Contentful Paint, and first-byte time on the largest images per template. Treat regressions in any of these as performance bugs, not aesthetics.

Teams building reading-heavy or test-prep sites can apply these patterns directly. The static-publishing patterns documented at pass4-sure.us and whats-your-iq.com follow exactly this stack.

What to Watch

The immediate developments worth tracking through 2026 and 2027:

  • Chrome's JPEG XL position. A Chrome shipment ends the dual-format era for the open web.
  • AVIF animation maturity. AVIF supports animated sequences, but tooling is still uneven. Expect this to converge.
  • Camera-side neural codecs. Apple, Sony, and Samsung have public research and patents in this space.
  • C2PA mandates. Several jurisdictions are considering provenance requirements for AI-generated and journalistic imagery.
  • Display-side gain map standards. The Ultra HDR pattern (base image plus gain map) is moving from Google-specific to a likely ISO standard.

Image formats reward patience. The decisions that matter most are which formats to use as canonical sources, what metadata to preserve, and which delivery pipelines to maintain. Browsers will catch up with whatever wins, eventually. The teams that prepared for AVIF in 2020 paid no migration cost in 2024. The teams preparing for JPEG XL and HDR today will pay no migration cost in 2027.

A Closing Note on Image Format Diligence

Across two decades of image format transitions, the teams that came out best were the ones that maintained two disciplines in parallel: lossless masters in well-supported formats, and automated derivative pipelines that could regenerate every public-facing asset from those masters. When WebP arrived, those teams produced WebP variants in an afternoon. When AVIF arrived, the same pipeline added AVIF in a similar afternoon. When JPEG XL eventually wins broad browser support, the same pattern will apply.

The teams that struggled were the ones whose canonical assets were already lossy, were stored across mixed proprietary tools, or had been edited and re-encoded so many times that the original quality was no longer recoverable. Their next format transition cost them months of recovery work or, worse, permanent quality loss they could not fix.

The lesson generalizes beyond images. Every category of file format has a parallel story: archival sources, automated derivatives, disciplined provenance, and patience while standards bodies finish their work. Image formats are simply the most visible case because the artifacts are visible to the user.

References

  1. ISO/IEC 18181-1:2022, JPEG XL Image Coding System, Part 1: Core coding system. https://www.iso.org/standard/77977.html
  2. AOMedia, AV1 Image File Format (AVIF) Specification 1.1.0. https://aomediacodec.github.io/av1-avif/
  3. Coalition for Content Provenance and Authenticity, Technical Specification 2.1. https://c2pa.org/specifications/specifications/2.1/
  4. Alakuijala, J., Sneyers, J., et al. JPEG XL next-generation image compression architecture and coding tools. SPIE Applications of Digital Image Processing XLII, 2019. https://doi.org/10.1117/12.2529237
  5. ITU-R BT.2100-2, Image parameter values for high dynamic range television. International Telecommunication Union, 2018. https://www.itu.int/rec/R-REC-BT.2100
  6. Rippel, O. and Bourdev, L. Real-time adaptive image compression. International Conference on Machine Learning, 2017. https://proceedings.mlr.press/v70/rippel17a.html
  7. Google Developers, Ultra HDR Image Format Specification 1.0. https://developer.android.com/guide/topics/media/platform/hdr-image-format
  8. W3C Working Draft, Media Source Extensions and HDR display modes, 2024. https://www.w3.org/TR/media-source-2/