zencodec

zencodec is the shared trait crate that defines the common API for all zen* image codecs.

zencodec contains no pixel encoding or decoding logic — that lives in the individual codec crates. It does include shared metadata parsing needed for nearly every image format: pixel-descriptor derivation from CICP/ICC metadata (with identification delegated to zenpixels::icc, covering 163 RGB + 18 grayscale web-corpus profiles), EXIF orientation extraction, ISO 21496-1 gain map parsing and serialization, and format detection via magic bytes. no_std compatible (requires alloc), forbid(unsafe_code).

Import as zencodec — zencodec::encode, zencodec::decode, and the shared types (metadata, limits, format detection, color emission) at the root.

Quick start

[dependencies]
zencodec = "0.1.26"

zencodec defines the traits; a concrete codec crate (here zenjpeg) supplies the pixels. The three-layer Config → Job → Encoder/Decoder shape is identical across every zen* codec, so swapping zenjpeg for zenwebp, zengif, zenavif, … changes only the config type:

use std::borrow::Cow;
use zenjpeg::{JpegEncoderConfig, JpegDecoderConfig};
use zencodec::encode::{EncoderConfig, EncodeJob, Encoder};
use zencodec::decode::{DecoderConfig, DecodeJob, Decode};

// Encode. `with_generic_quality` is the codec-agnostic knob on a calibrated
// 0.0..=100.0 scale (NOT 0.0..=1.0); higher is better. Read it back with
// `generic_quality()` — `None` means the codec has no quality dial.
let config = JpegEncoderConfig::new().with_generic_quality(85.0);
// (assuming pixels: PixelSlice from your pipeline)
let output = config.job().encoder()?.encode(pixels.as_slice())?;
let jpeg_bytes = output.into_vec();

// Decode
let config = JpegDecoderConfig::new();
let decoded = config.job().decoder(Cow::Borrowed(&jpeg_bytes), &[])?.decode()?;
let pixels = decoded.into_buffer();

For untrusted input, attach a resource limit and a cancellation token to the job — see Untrusted input below.

Crates in the zen* family

Crate	Format	Repo
zenjpeg	JPEG	imazen/zenjpeg
zenwebp	WebP	imazen/zenwebp
zenpng	PNG	imazen/zenpng
zengif	GIF	imazen/zengif
zenavif	AVIF	imazen/zenavif
zenjxl	JPEG XL	imazen/zenjxl
zenbitmaps	PNM/BMP/Farbfeld	imazen/zenbitmaps
heic	HEIC/HEIF	imazen/heic
zentiff	TIFF (experimental)	imazen/zentiff
zenpdf	PDF (experimental)	imazen/zenpdf

Architecture

Every codec follows a three-layer pattern:

Config     →  reusable, Clone + Send + Sync, 'static — consumed by job()
Job        →  per-operation, owns config + stop token + limits + metadata
Executor   →  borrows pixel data or file bytes, consumes self to produce output

ENCODE:  EncoderConfig → EncodeJob → Encoder / AnimationFrameEncoder
DECODE:  DecoderConfig → DecodeJob<'a> → Decode / StreamingDecode / AnimationFrameDecoder

Config lives in a struct and gets shared across threads. A web server keeps one JpegEncoderConfig at quality 85 for all requests and clones it per-request. Calling job() consumes the config — clone first if you need it again. Job owns its config, cancellation token, resource limits, and metadata. Executor borrows pixels or bytes and consumes itself to produce output.

Each layer also has object-safe Dyn* variants for codec-agnostic dispatch:

DynEncoderConfig → DynEncodeJob → DynEncoder / DynAnimationFrameEncoder
DynDecoderConfig → DynDecodeJob → DynDecoder / DynStreamingDecoder / DynAnimationFrameDecoder

Blanket impls generate the dyn API automatically — codec authors implement the generic traits and get dyn dispatch for free.

Untrusted input: limits, cancellation, errors

Two server-critical knobs — ResourceLimits and a cancellation StopToken — are attached to the job, not the config: DecodeJob/EncodeJob expose .with_limits(limits) and .with_stop(token) (both consume-and-return-self builders), so the per-request job carries them while the shared config stays immutable. Everything below is on the root: use zencodec::{ResourceLimits, StopToken};.

Constructing ResourceLimits

ResourceLimits is a plain struct with pub Option<_> fields — build it three ways:

use zencodec::ResourceLimits;

// 1. Server preset. `for_untrusted_input()` fills generous-but-bounded caps:
//    max_pixels 120 MP/frame, max_total_pixels 200 MP (all frames),
//    max_width/max_height 16384 each, max_memory_bytes 1 GiB,
//    max_input_bytes 256 MiB, max_frames 65 536, max_animation_ms 1 hour.
//    (`Default`/`none()` is the OPPOSITE — every field `None`, i.e. UNLIMITED;
//    use that only for trusted input.) Tighten any field with a `with_*` builder:
let limits = ResourceLimits::for_untrusted_input()
    .with_max_pixels(4_000_000)          // 4 MP cap for a thumbnail service
    .with_max_memory(256 * 1024 * 1024); // 256 MiB

// 2. From scratch off the unlimited default, set only what you need:
let limits = ResourceLimits::none()
    .with_max_pixels(16_000_000)         // pixels = width × height (per frame)
    .with_max_input_bytes(8 * 1024 * 1024); // bytes of encoded input (decode)

// 3. Direct field set (every field is public):
let mut limits = ResourceLimits::default();
limits.max_width = Some(8192);           // pixels
limits.max_height = Some(8192);          // pixels

Units: max_pixels / max_total_pixels are pixel counts (width × height, the latter ×frame_count); max_width / max_height are pixels; max_memory_bytes, max_input_bytes, max_output_bytes are bytes; max_frames a count; max_animation_ms milliseconds. A None field means that dimension is unchecked.

ResourceLimits also carries an allocation-fallibility preference, prefer_fallible_allocations (an AllocPreference: CodecDefault / Fallible / Infallible). CodecDefault (the default) lets each codec choose — decoders favour the fallible try_reserve path on untrusted input, encoders the faster infallible vec!; for_untrusted_input() presets Fallible. Override with .with_prefer_fallible_allocations(AllocPreference::Fallible).

Constructing a cancellation StopToken

StopToken is re-exported from the almost-enough crate (cargo add almost-enough). Make a Stopper (cheap, Clone, 8 bytes), erase it into a StopToken, hold a clone of the Stopper to fire later from any thread:

use zencodec::StopToken;

let stopper = almost_enough::Stopper::new();
let token = StopToken::new(stopper.clone()); // or: stopper.clone().into()

// Fire from a deadline / client-disconnect watcher — `cancel()` signals every clone:
std::thread::spawn({
    let stopper = stopper.clone();
    move || { /* on timeout or disconnect: */ stopper.cancel(); }
});
// (For a no-op token when you don't need cancellation, use `zencodec::Unstoppable`.)

End-to-end: decode untrusted bytes with a limit + a stop token

.with_limits() and .with_stop() chain onto the job before you ask for the decoder. probe() (also on the job) is an O(header) parse — validate against the limits before the codec allocates pixels, then attach both to the real decode:

use std::borrow::Cow;
use zencodec::{ResourceLimits, StopToken, CodecErrorExt};
use zencodec::decode::{DecoderConfig, DecodeJob, Decode};

let limits = ResourceLimits::for_untrusted_input().with_max_memory(256 * 1024 * 1024);
let stopper = almost_enough::Stopper::new();
let token = StopToken::new(stopper.clone());

// 1. Probe the header and reject oversized inputs before any pixel allocation.
let info = config.clone().job().probe(bytes)?;
limits.check_image_info(&info)?; // -> Err(LimitExceeded) if too big

// 2. Attach BOTH to the job, then decode. Order is free; `decoder()` consumes the job.
let decoded = config.job()
    .with_limits(limits)
    .with_stop(token)
    .decoder(Cow::Borrowed(bytes), &[])? // &[] = native pixel format
    .decode()?;
let pixels = decoded.into_buffer();

.with_stop() is honored only by codecs whose DecodeCapabilities::stop() is true; on others it is a silent no-op (the decode still completes correctly, just not interruptibly). check_image_info only sees what the header reports — keep the max_memory_bytes cap so a codec that under-reports still can't over-allocate.

Classifying a codec's error for an HTTP response

A codec surfaces errors one of two ways; both let you route on a coarse ErrorCategory instead of naming the concrete enum:

• Native enum + CategorizedError. The codec keeps its own error type and maps every variant to one category. e.category() classifies it on the typed path, and a located whereat::At<E> keeps the inner error's category (the impl forwards through At).
• The shared CodecError envelope. The codec sets type Error = whereat::At<CodecError>. Because that is a single concrete type, the category — and the originating codec name ("zenjpeg", so you can tell codecs apart) — survive type erasure: after dyn dispatch hands back a Box<dyn Error> — or you map into anyhow — e.error_category() / e.codec_error() recover them by downcast (None only when the error isn't from a zen codec). A native enum can't be recovered once erased — the erased value is a dyn Error, not a dyn CategorizedError — so the envelope is what a codec-agnostic pipeline reaches for. CodecError can also be used with no detail at all (CodecError::new(Some("zenpng"), ErrorCategory::Image(ImageError::Malformed))) by a codec that has no error enum of its own. At is re-exported (zencodec::At<zencodec::CodecError>), so adopting it needs no direct whereat dependency.

use zencodec::enough::StopReason;
use zencodec::{
    CategorizedError, CodecErrorExt, ErrorCategory, ImageError, InvalidKind, RequestError,
    ResourceError,
};

// Typed path shown here via `category()`; on a `Box<dyn Error>` from dyn dispatch
// use `e.error_category()` and match `Some(..)` / `None` instead. The origin-first
// shape lets you route on the outer arm (`Image` / `Request` / `Resource` /
// `Stopped`) and only destructure when a sub-kind changes the answer.
let http = match config.job().with_stop(token).decoder(Cow::Borrowed(bytes), &[]) {
    Ok(_decoder) => { /* _decoder.decode()? */ 200 }
    Err(e) => match e.category() {
        // Stopped via its Stop token.
        ErrorCategory::Stopped(StopReason::Cancelled) => 499, // client went away
        ErrorCategory::Stopped(StopReason::TimedOut)  => 504,
        // A configured resource cap was hit.
        ErrorCategory::Resource(ResourceError::Limits(_)) => 413,
        // "Can't process this" — the bytes need a different codec, or the request
        // asked for an op / pixel format / CMS transform this codec doesn't do.
        ErrorCategory::Image(ImageError::Unsupported(_))
        | ErrorCategory::Request(RequestError::Unsupported(_))
        | ErrorCategory::Request(RequestError::CmsRequired) => 415,
        // Valid input, but a policy declined it — Decode vs Encode only matters
        // if you want to word the message differently; both are 422 here.
        ErrorCategory::Policy(_) => 422,
        // Bad client-supplied bytes (malformed / truncated), or a bad request.
        ErrorCategory::Image(_) => 400,
        ErrorCategory::Request(RequestError::Invalid(
            InvalidKind::Parameters | InvalidKind::Buffer,
        )) => 400,
        // Server-side: API misuse (out-of-sequence call), OOM, I/O, or a bug.
        // `Internal(Bug)` vs `Internal(Dependency)` both still read as 500 for
        // HTTP purposes; the split pays off in telemetry, not routing.
        ErrorCategory::Request(RequestError::Invalid(InvalidKind::State))
        | ErrorCategory::Resource(ResourceError::OutOfMemory)
        | ErrorCategory::Io(_)
        | ErrorCategory::Internal(_) => 500,
        _ => 500, // ErrorCategory is #[non_exhaustive]
    },
};

ErrorCategory is the coarse routing axis (it's Copy, and LimitsExceeded(LimitKind) even tells you which cap). When you need more detail than the category, the typed extractors on CodecErrorExt recover the cause from the source() chain — e.limit_exceeded() for the exact limit values, e.unsupported_operation() for which operation, or e.find_cause::<T>() for a codec-specific type. category() is available once a codec implements CategorizedError; the zencodec cause types (LimitExceeded, UnsupportedOperation, enough::StopReason) implement it, so a codec's mapping is usually a one-line delegation per variant. CodecErrorExt::error_category() is the same answer recovered through erasure — it works on any Box<dyn Error> holding a CodecError (bare or At-wrapped), so the routing survives dyn dispatch and anyhow.

Controlling decode parallelism

Zen codecs parallelize with rayon's ambient pool. The responsibility splits cleanly: the codec chooses sequential vs parallel (ResourceLimits::threading → ThreadingPolicy::{Sequential, Parallel}); the caller chooses the thread count by sizing a pool and running the decode inside rayon::ThreadPool::install — you don't ask the codec to cap threads from the inside.

DynDecoder (one-shot) is intentionally not Send — it may borrow your input zero-copy. That is not an obstacle to capping threads: decode() returns an owned, Send DecodeOutput, so construct and consume the decoder inside the closure and only the result crosses back out. The codec's internal rayon work then runs on your sized pool:

use std::borrow::Cow;
use rayon::ThreadPoolBuilder;
use zencodec::decode::{DecodeOutput, DynDecoderConfig};

let pool = ThreadPoolBuilder::new().num_threads(2).build()?;
let out: DecodeOutput = pool.install(|| {
    config.dyn_job()
        .into_decoder(Cow::Borrowed(&bytes), &[])?   // Box<dyn DynDecoder> — not Send, but local
        .decode()                                     // -> DecodeOutput (Send), consumes the decoder
})?;
// the non-Send decoder lived and died on a pool worker; only owned pixels came back

Run many decodes under one capped pool the same way: pool.install(|| inputs.par_iter().map(decode_one).collect()), each decoder built inside its own task. Need a live decoder on another thread (not just thread-capping)? Use the streaming path: DynStreamingDecoder is Send by contract (it owns/copies its data), so it can move across a thread boundary — one-shot trades that for zero-copy borrowing.

Exception — native-threaded codecs. AV1 (rav1d / zenrav1e / AVIF) spawns OS threads, not rayon, so install() has no effect on them; cap those with codec-specific config (e.g. AvifEncoderConfig::with_threads(4)). See ThreadingPolicy for the full model.

These patterns are exercised end-to-end in zencodec-testkit/tests/decode_parallelism.rs.

Encode fidelity

EncoderConfig::with_fidelity is the codec-agnostic way to ask for a quality level — infallible and best-effort: the codec does what it can and substitutes the rest.

use zencodec::encode::{EncoderConfig, Fidelity};

let cfg = my_encoder_config
    .with_fidelity(Fidelity::Lossless);              // mathematically exact
//  .with_fidelity(Fidelity::ssim2(90.0))            // aim at SSIMULACRA2 ~= 90
//  .with_fidelity(Fidelity::butteraugli(1.0))       // aim at butteraugli max-norm ~= 1.0
//  .with_fidelity(Fidelity::zensim_b(90.0))         // aim at zensim (ZensimProfile::B) ~= 90
//  .with_fidelity(Fidelity::codec_quality(85.0));   // the codec's own native dial

Fidelity is either Lossless or Lossy(LossyTarget), where a LossyTarget is:

• ApproxSsim2(score) — a one-shot SSIMULACRA2 target (a real cross-codec metric).
• ApproxButteraugli(distance) — a one-shot butteraugli max-norm distance.
• ApproxZensimB(score) — a one-shot zensim target, pinned to ZensimProfile::B.
• CodecSpecificQuality(q) — the codec's own native quality scale (meaning differs per codec).

These are blind, single-pass: the target maps to a native dial in one encode, no re-encode loop. A codec that hasn't implemented native fidelity still behaves sensibly — the default bridges to the legacy with_lossless / with_generic_quality setters. Read back what the codec resolved to with resolved_target_fidelity() -> Option<Fidelity>.

A LosslessMode container variant (lossless coding within a loss budget — the screen-content path) and a fail-fast try_with_fidelity verdict are designed but deferred while their semantics settle; see the reserved blocks in src/fidelity.rs and imazen/zencodec#104.

Key Design Decisions

Color management is not the codec's job. Decoders return native pixels with ICC/CICP metadata. Encoders accept pixels as-is and embed the provided metadata. The caller handles CMS transforms.

Format negotiation over conversion. Decoders take a ranked &[PixelDescriptor] preference list and pick the first they can produce without lossy conversion. Pass &[] for native format.

Capabilities over try/catch. Codecs declare their capabilities as const EncodeCapabilities / DecodeCapabilities structs. Check before calling instead of catching UnsupportedOperation errors.

Pixel types from zenpixels. All pixel interchange types (PixelSlice, PixelBuffer, PixelDescriptor, etc.) are defined in the zenpixels crate. All zen* crates depend on zenpixels directly.

Metadata Retention

Re-encode and recompress pipelines need to decide what metadata survives. Metadata::filtered applies a MetadataPolicy, so callers never hand-parse EXIF:

use zencodec::{MetadataPolicy, MetadataFields, IccRetention, exif::{ExifPolicy, Retention}};

// Decode → filter → re-encode. `Web` (recommended for publishing) keeps the ICC profile
// (unless a redundant sRGB), EXIF orientation + rights, and CICP/HDR color
// signaling — and strips GPS, timestamps, camera info, thumbnail, and XMP.
let kept = decoded_meta.filtered(&MetadataPolicy::Web);

// Presets: PreserveExact (keep all, incl. duplicate sRGB), Preserve (drop dup
// sRGB), Web, ColorAndRotation (only what places pixels), Custom.
let minimal = decoded_meta.filtered(&MetadataPolicy::ColorAndRotation);

// Per-field control — drop only the thumbnail, keep everything else:
let policy = MetadataPolicy::Custom(
    MetadataFields::KEEP_ALL.with_exif(ExifPolicy::KEEP_ALL.with_thumbnail(Retention::Discard)),
);
let no_thumb = decoded_meta.filtered(&policy);

MetadataFields encapsulates EXIF in an ExifPolicy with seven keep/discard categories — orientation, rights, thumbnail, gps, datetimes, camera, other — and three-way ICC handling (IccRetention::{Drop, KeepNonSrgb, Keep}). EXIF passes through byte-unchanged (zero-copy) when no category is dropped, and is rewritten — offsets recomputed — only when pruning. CICP/HDR are color signaling (dropping them changes displayed pixels), so the presets keep them; a Custom policy can drop them. The structured parser/editor is public as zencodec::exif::Exif (parse → filtered/edit → to_bytes) for direct EXIF work — including setting Copyright/Artist (set_copyright / set_artist, with a TextEncoding choice of Exif 2.x ASCII or Exif 3.0 UTF-8) and Orientation (set_orientation, insert-or-replace).

Privacy is an explicit choice — enforced at compile time. Retention is a transient decision made when you hand metadata to the encoder, not a field stored on Metadata. The blessed path is job.with_metadata_policy(meta, MetadataPolicy::Web) (privacy-safe: strips camera/GPS, keeps orientation + rights) or PreserveExact (verbatim). The old unguarded with_metadata(meta) still works but is #[deprecated] — the compiler warns at every call site that picks no policy, so you can't propagate metadata without choosing retention by accident. It's a compile-time nudge, not a semver break: existing code keeps compiling, but the warning points you at the safe call. The filter runs before the codec sees the record, so a codec only ever receives exactly what the policy kept. The carried bytes stay untouched until then, so you can still pull metadata.exif out, edit it with any EXIF library, and put it back via with_exif.

To stamp rights in one line — Metadata::none().with_copyright("© 2026 You") builds (or merges into) the EXIF blob (ASCII); or build it directly with Exif::new(TextEncoding::Ascii).set_copyright(…) → to_bytes() — Exif::new requires the Exif 2.x-vs-3.0 field-type choice (type 129 is read by almost nothing, so it's never a silent default).

Metadata retention, color emission, and orientation are the three correctness signals an encode has to get right; docs/correctness-model.md describes how the framework resolves each one before the codec runs so a codec can't quietly clobber it. The zencodec-testkit crate verifies a codec honors that contract — check_metadata_no_leak re-parses the embedded EXIF to prove a policy's drops actually happened, and check_cross_path_pixel_equivalence diffs every feeding mode.

Color Emission

The encode-side dual of color resolution: which color carriers (ICC vs CICP) should an encode write? resolve_color_emit decides — a pure, no_std, CMS-free function of the source color, the target's carrier capabilities, and a policy:

use zencodec::{resolve_color_emit, ColorEmitPolicy, IccDisposition};

let plan = resolve_color_emit(&source_color, &target_caps, ColorEmitPolicy::Balanced);
// plan.cicp: Option<Cicp>   — write this CICP (JXL/AVIF/HEIC nclx, PNG cICP) if the format carries it
// plan.icc:  IccDisposition — KeepSource | SynthesizeFrom(Cicp) | Drop

ColorEmitPolicy picks the tradeoff: Compatibility (widest reader support), Balanced (default — CICP where it's a spec-mandated safe sole carrier, an ICC companion otherwise), Compact (smallest — prefer CICP, drop the ICC), Verbatim (carry the source's signals unchanged), or Custom(ColorEmitFields). A target advertises its carriers via EncodeCapabilities::{cicp_is_valid_carrier, cicp_safe_sole_carrier}. The plan never emits a redundant SynthesizeFrom(sRGB); a codec lowers a SynthesizeFrom through zenpixels-convert's transfer-aware synthesize_icc_for_cicp (a bundled const profile, or — with its cms-moxcms feature — a generated one) so an HDR transfer is never mis-tagged with an SDR profile and color is never silently dropped. The names carry the emit direction so they can't be confused with the decode-side SourceColor. Design + rejected alternatives: docs/color-emit-model.md.

What's in this crate

Module	Contents
zencodec::encode	EncoderConfig, EncodeJob, Encoder, AnimationFrameEncoder, EncodeOutput, EncodeCapabilities, EncodePolicy, best_encode_format, dyn dispatch traits (DynEncoderConfig, DynEncodeJob, DynEncoder, DynAnimationFrameEncoder)
zencodec::decode	DecoderConfig, DecodeJob, Decode, StreamingDecode, AnimationFrameDecoder, DecodeOutput, DecodeCapabilities, DecodePolicy, DecodeRowSink, SinkError, OutputInfo, SourceEncodingDetails, negotiate_pixel_format, is_format_available, dyn dispatch traits (DynDecoderConfig, DynDecodeJob, DynDecoder, DynStreamingDecoder, DynAnimationFrameDecoder)
zencodec::estimate	ResourceEstimate (predicted peak memory / wall + CPU time / core-scaling, all Option), ComputeEnvironment (cores, RAM, SimdTier), ImageCharacteristics, SimdTier, ThreadingInformation — codec-agnostic resource estimation, surfaced via EncoderConfig::estimate_encode_resources / DecoderConfig::estimate_decode_resources
zencodec::gainmap	GainMapInfo, GainMapParams, GainMapChannel, GainMapDirection, GainMapPresence, Iso21496Format (wire-format variant: AvifTmap, JxlJhgm, JpegApp2BodyWithUrn; the original JpegApp2 is deprecated since 0.1.20), ISO_21496_1_URN, ISO_21496_1_PRIMARY_APP2_BODY, serialize_iso21496_fmt / serialize_iso21496_fmt_into / parse_iso21496_fmt, GainMapParseError — cross-codec gain map types and wire-format helpers (ISO 21496-1)
zencodec::exif	Structured EXIF/TIFF: Exif (borrowing parse → prune → serialize), ExifPolicy (7 keep/discard categories), Retention, ByteOrder, retain
zencodec::helpers	Codec implementation helpers (not consumer API) — shared boilerplate for trait implementors, plus the lightweight parse_exif_orientation accessor
root	ImageFormat, ImageFormatDefinition, ImageFormatRegistry (format detection via ImageFormatRegistry::detect()), ImageInfo, Metadata, MetadataPolicy, MetadataFields, IccRetention, Exif, ExifPolicy, Retention, ByteOrder, Orientation, OrientationHint, ResourceLimits, AllocPreference, LimitExceeded, ThreadingPolicy, UnsupportedOperation, CodecErrorExt, find_cause, Unsupported, Extensions, AnimationFrame, OwnedAnimationFrame, resolve_color_emit, ColorEmitPolicy, ColorEmitPlan, ColorEmitFields, IccDisposition, CicpEmission, ColorAuthority, Cicp, ContentLightLevel, MasteringDisplay, StopToken, Unstoppable

zencodec has no feature flags. The full API is always available.

Limitations

• Contains no codec logic — traits, types, and format detection only.
• ImageFormat enum is not extensible at runtime (the Custom variant requires a &'static definition).
• Always no_std + alloc (no std feature gate).

MSRV

Rust 1.88+, 2024 edition.

License

Licensed under either of Apache-2.0 or MIT at your option.

Image tech I maintain

Codecs ¹	zenjpeg · zenpng · zenwebp · zengif · zenavif · zenjxl · zenbitmaps · heic · zentiff · zenpdf · zensvg · zenjp2 · zenraw · ultrahdr
Codec internals	zenjxl-decoder · jxl-encoder · zenrav1e · rav1d-safe · zenavif-parse · zenavif-serialize
Compression	zenflate · zenzop · zenzstd
Processing	zenresize · zenquant · zenblend · zenfilters · zensally · zentone
Pixels & color	zenpixels · zenpixels-convert · linear-srgb · garb
Pipeline & framework	zenpipe · zencodec · zencodecs · zenlayout · zennode · zenwasm · zentract
Metrics	zensim · fast-ssim2 · butteraugli · zenmetrics · resamplescope-rs
Pickers & ML	zenanalyze · zenpredict · zenpicker
Products	Imageflow image engine (.NET · Node · Go) · Imageflow Server · ImageResizer (C#)

_{¹ pure-Rust, #![forbid(unsafe_code)] codecs, as of 2026}

General Rust awesomeness

zenbench · archmage · magetypes · enough · whereat · cargo-copter

Open source · @imazen · @lilith · lib.rs/~lilith