oximg

High-performance image compression in Rust: a library, a CLI, and a self-hostable HTTP server (PoC). JPEG, PNG, WebP — and AVIF with the avif feature — in and out; sources are format-sniffed by magic bytes and re-encoded in their own format. On imgproxy's official benchmark harness, run on the same AWS instance types as their published results, oximg leads every format cell on both x86-64 and Graviton while resizing in linear light at measurably higher output quality (see Benchmarks).

Features

• HTTP resize service: GET /resize/{w}/{h}/{file} fits the source within w x h (never enlarges) and re-encodes it in its own format. 0 leaves an axis unconstrained (/resize/750/0/… is width-only — what srcset w descriptors and Next.js loaders emit), and {file} may span directories, so S3-style prefixes and nested trees are addressable as-is. Optional imgproxy-style HMAC URL signing.
• Cloudflare Images URL compatibility: mount a second route (OXIMG_OPTIONS_PREFIX) speaking the option-list grammar — /image/width=750,quality=80/path/to/photo.png — so URLs built for Cloudflare Images survive a migration without a rewrite layer, per-request quality included.
• Sources: a local directory, any HTTP(S) origin, or a private GCS bucket (gs:// with GCP-attached credentials — no public bucket, no public-endpoint egress). Streaming decode overlaps the download; transient fetch failures are retried, so a network blip is a slower response, not a broken image.
• Production operability: graceful SIGTERM drain, upstream fetch deadlines (slow-origin 504s distinct from broken-origin 502s), and an opt-in Prometheus /metrics page whose queue-wait/processing split tells "needs more CPU" apart from "sources got bigger".
• Quality-first processing: resizing happens in linear light on 16-bit samples with Lanczos3, JPEG sources are decoded supersampled (DCT shrink-on-load kept ≥ 1.7x the target), and alpha is premultiplied across the resample — the properties behind the SSIMULACRA2 scores in Benchmarks.
• Performance as architecture, not flags: per-arch row-streaming SIMD resize kernels (AVX2 on x86-64, NEON on aarch64, both verified against an f64 reference), JPEG decode fused with resize+encode on a second thread under low load, request coalescing for concurrent identical URLs (per-process — a horizontally scaled deployment gets its dedup from the CDN in front, not from here), and CPU concurrency pinned to the core count. Peak memory stays at a fraction of imgproxy's under identical load (BENCH.md).
• Tunable profiles: the default maximizes quality per byte (progressive jpegli); one env flip (OXIMG_JPEG_PROGRESSIVE=0) trades ~10% output size for the lowest latency at unchanged pixels. PRESET=fast|small selects mozjpeg profiles instead.
• Self-contained deploys: multi-arch Docker images (linux/amd64 + linux/arm64) on Docker Hub (oximg/oximg) and GHCR (ghcr.io/oximg/oximg); a single static-leaning binary otherwise.

Supported formats

Sources are identified by magic bytes (extensions are never trusted). By default the output format is the source's own; any decode column combines with any encode column:

Cross-format output: append an imgproxy-style @{fmt} token to the filename — /resize/300/200/photo.jpg@webp (jpg/jpeg, png, webp, avif; jxl is reserved). Only exact tokens count, so photo@2x.jpg is still a filename. Precedence: explicit @{fmt} > Accept negotiation > source format. Negotiation is opt-in: set OXIMG_AUTO_FORMAT to a preference list (e.g. avif,webp) and bare-URL responses follow the request's Accept header; every response then carries Vary: Accept (make sure your CDN honors it or normalizes Accept into the cache key — explicit @{fmt} URLs avoid the issue entirely, which is what signed deployments should prefer since headers are outside the signature). Alpha sources encoded to JPEG are flattened in linear light onto OXIMG_FLATTEN_BG (hex RRGGBB, default white). Encode settings are keyed by the output format, using the same knobs as same-format requests.

Choose the preference order by your goal: the AVIF defaults target fidelity, not minimum bytes — at default quality settings AVIF output measures 10–28% larger than WebP on photographic sources, and costs the more expensive encode. If the deployment's goal is byte reduction, prefer webp,avif (or webp alone), or lower OXIMG_AVIF_QUALITY until AVIF earns its slot; put avif first only after comparing sizes on your own corpus at your own settings. Also note what negotiation does not cover: when it doesn't fire (client sends Accept: */* — link-preview scrapers, social-card fetchers, curl integrations), the source format is kept, and PNG output defaults to lossless RGB(A) — a large photographic PNG stays large unless OXIMG_PNG_QUANTIZE=1 is set. Deployments that care about those clients should enable quantization or prefer explicit @{fmt} URLs over relying on negotiation. On flat graphics (charts, screenshots, text-heavy panels), a quantized PNG is often both smaller and truer to the source than any WebP quality setting — worth remembering when tuning OXIMG_AUTO_FORMAT for mixed content.

Orientation: every source format auto-rotates — JPEG EXIF, PNG eXIf, WebP EXIF chunks, and AVIF irot/imir transforms. The target box applies to the displayed frame and the pixels come out upright in every output format (the metadata itself is not forwarded, so nothing double-rotates). OXIMG_AUTO_ROTATE=0 restores the raw stored orientation.

ICC profiles: a source's color profile (JPEG APP2 chain, PNG iCCP, WebP ICCP, AVIF colr) passes through byte-for-byte into any output format, across format conversion included. RGB pixels are never color-converted. This matters for wide-gamut sources: the common proxy default is to normalize pixels to sRGB and strip the profile, which permanently clips every color outside the sRGB gamut — a Display P3 phone photo loses exactly the saturated reds and greens that made it worth shooting in P3. oximg keeps the pixels and the profile as they were, so wide-gamut images render on a wide-gamut display the way the original did (and identically everywhere else). OXIMG_ICC=0 opts into stripping instead.

CMYK/YCCK JPEG sources (print-workflow assets) are the one exception, since no browser renders CMYK pixels: they are converted to sRGB — through the embedded CMYK profile (moxcms, relative colorimetric, like imgproxy/libvips) when one is present, with the naive composite browsers use otherwise — and the CMYK profile is consumed, never passed through. OXIMG_ICC=0 skips profile extraction entirely, so it also selects the naive conversion.

Pipeline

source bytes (local file or HTTP origin)
  → format sniff → decode
      JPEG: mozjpeg streaming decode, DCT shrink-on-load (kept ≥ 1.7x target size)
      PNG:  png crate (palette/gray/16-bit normalized to RGB(A)8)
      WebP: libwebp
      AVIF: dav1d (8/10/12-bit, all subsamplings, alpha, bilinear chroma upsampling)
  → linear-light resize: sRGB u8 → linear u16 → Lanczos3 → sRGB u8
      (alpha is premultiplied before resampling, unpremultiplied after;
       JPEG rows stream through in-tree ring-scheduled f32 row kernels —
       AVX2 on x86-64, NEON on aarch64, both verified against an f64
       reference — optionally fused with the decode on a second thread;
       other formats resize full-frame: pic-scale on x86-64, the same
       in-tree kernel on aarch64)
  → encode in the source format
      JPEG: jpegli, progressive (PRESET=fast / PRESET=small select mozjpeg profiles)
      PNG:  png crate | WebP: libwebp | AVIF: SVT-AV1 (10-bit 4:2:0, tune=ssim)

Concurrent identical requests are coalesced and share one result. CPU concurrency is pinned to the core count with a semaphore; the HTTP layer (axum/tokio) only does queueing and IO.

Benchmarks

imgproxy's official harness (DIV2K corpus over nginx, fit into 512x512, k6, all defaults) on the AWS instance types behind imgproxy's published numbers — req/s, higher is better, p95 in parentheses:

Cross-format cells (our harness extension; JPEG sources, oximg vs imgproxy):

At the same time, output quality is higher, not traded away: end-to-end JPEG at the same q80 scores +6 to +18 SSIMULACRA2 over imgproxy (77.5 vs 71.2 on the Kodak corpus, the gap widening with source size — and imgproxy at q90 with twice the bytes still scores lower), pure resize quality (lossless PNG path) scores 97.6 vs 81.9, and the AVIF default produces smaller files than imgproxy's default at +6.7 SSIMULACRA2.

• BENCH.md — full methodology and tables: official harness (local and AWS), sustained-load and memory measurements, presets.
• bench/quality/QUALITY.md — output quality (SSIMULACRA2) at matched settings vs imgproxy and sharp.

Install

Docker (recommended — multi-arch linux/amd64 + linux/arm64, AVIF included; both registries rebuild on every main push, so pin a version tag in production):

docker run -p 8081:8081 -v $PWD/images:/images:ro ghcr.io/oximg/oximg:latest
# or: docker.io/oximg/oximg:latest
curl "localhost:8081/resize/500/500/photo.jpg" -o out.jpg

Prebuilt binaries (GitHub Releases, v0.6.0+; Linux x86_64/aarch64 and macOS arm64; JPEG/PNG/WebP, no AVIF) — suited to CI asset pipelines where a Docker pull or a source build is too slow. Assets are oximg-<tag>-<target>.tar.gz with a .sha256 alongside; each is smoke-tested before upload. Linux builds link glibc >= 2.39 with libstdc++ static.

Homebrew (builds the latest release from source; JPEG/PNG/WebP):

brew install oximg/tap/oximg

Cargo (crates.io; add --features avif if SVT-AV1 >= 4.1 and dav1d are installed and visible to pkg-config):

cargo install oximg

From source (the Docker build needs no system dependencies — it compiles a pinned SVT-AV1 itself):

cargo build --release                    # JPEG, PNG, WebP
cargo build --release --features avif    # + AVIF (needs SVT-AV1 >= 4.1, dav1d)
IMAGES_DIR=./images PORT=8081 ./target/release/oximg   # = oximg serve

Release channels lag main: crates.io and the brew formula ship the last tagged release, while the Docker images rebuild on every main push. The npm package @oximg/oximg is a name reservation that points here.

Serving

URL grammars. The positional route is /resize/{w}/{h}/{file}[@fmt]; 0 leaves an axis unconstrained, and {file} may span directories. Setting OXIMG_OPTIONS_PREFIX mounts a second route speaking the Cloudflare Images option grammar at that prefix:

/image/width=750,quality=80/albums/2026/photo.png

with width/height (1-8192; one suffices, the other axis follows the aspect ratio), quality (1-100, applied to whichever format the output resolves to; PNG output is lossless and ignores it), and format (jpeg|png|webp|avif, or auto = the same Accept negotiation as a bare positional URL, which also runs when format is absent). Unknown or duplicate options answer 400 naming the key — Cloudflare silently ignores unknown options, but a silently dropped fit=cover changes the output, so the divergence is deliberate. The filename is taken literally on this route (no @fmt token).

Sources. With OXIMG_SOURCE_BASE_URL unset, sources come from IMAGES_DIR. Set it and the scheme selects the transport:

• https://host/prefix — anonymous HTTP. Exposure prerequisite: no credentials are sent, so the origin must be anonymously readable; for an object-store bucket that means public objects, and anyone who can guess a path can fetch the original at full resolution, bypassing every resize/signing/CDN control in front.
• gs://bucket[/prefix] — a private GCS bucket, read directly with GCP-attached credentials (GKE Workload Identity, Cloud Run, and GCE metadata credentials; tokens cached and refreshed; boot fails closed with a clear message when no credentials are reachable). service_account JSON keys are not supported — on GCP use Workload Identity, off GCP use the HTTP mode. s3:// is planned (issue #11).

Streaming decode overlaps the download in every mode. Connection-level transients (reset, refused, DNS blips) are retried once before any body bytes are consumed, and the gs:// mode also retries 429/5xx SDK-style; oximg_upstream_retries_total counts both.

Error classes follow fault, not convenience: a source key that no store can serve — past an object store's key-length limit, or refused by the origin as a malformed request (400/414) — answers 400, and an absent object 404. Only a genuinely unwell upstream (connect failure, reset, 5xx) answers 502, with slow origins split off as 504. This matters downstream: CDNs retry and fail over on 5xx but pass 4xx through to their error cache, so misfiling a client error as an upstream failure both inflates the 5xx rate an operator watches and turns a crawler into origin load. oximg_upstream_fetch_total splits the same way (rejected and not_found apart from error), so that series stays a signal of upstream health. Over-length keys are refused locally, without a round trip.

Source paths are validated component-wise — ./.. components, empty components, \, ?, #, and control bytes answer 400. Local sources also pass a symlink-containment check (a path resolving outside IMAGES_DIR answers 404), and remote paths are re-encoded segment-wise so a percent in a name is never double-decoded upstream.

URL signing (optional): set OXIMG_KEY and OXIMG_SALT (hex) to require imgproxy-style signed URLs — /{base64url(HMAC-SHA256(key, salt || path))}/resize/{w}/{h}/{file}, and the same scheme over {prefix}/{options}/{file} on the options route. The signed path is the percent-decoded form, so one signature covers every URL encoding of the same source.

Graceful shutdown: on SIGTERM (what docker stop, Kubernetes, and Cloud Run send) or SIGINT the server stops accepting connections, finishes in-flight requests, and exits 0. There is no drain timeout of its own — the orchestrator's grace period backstops a response that never finishes, so allow a few seconds more than your slowest expected encode.

CLI

One-shot commands over the same pipeline, no server:

oximg resize photo.jpg 1600 1600 out.webp     # fit within 1600x1600; format from the extension
oximg resize photo.jpg 800 800 out.jpg -q 70  # JPEG quality 70 (--preset fast|small for mozjpeg)
oximg resize photo.jpg 750 0 out.jpg          # width-only: height follows the aspect ratio
oximg resize photo.jpg 0 0 out.webp           # 0 0 = re-encode at the source's own size
oximg probe photo.webp                        # format + stored dimensions, header-only

The output format is -f/--format, else the <out> extension, else the source's own format — the same precedence idea as the server's @fmt grammar. The OXIMG_* encode knobs below apply to CLI encodes the same way. Usage errors exit 2; processing failures exit 1.

Library

The oximg::pipeline module is usable without the HTTP server — process/process_path take a Params and return the re-encoded bytes plus their format, probe reads just the header. Depend on it with default-features = false to drop the entire HTTP stack (axum, tokio, ureq, hmac, sha2, serde_json); add features = ["avif"] for AVIF. process_url and process_gcs (remote sources) need the server feature.

Failures are typed: every entry point returns pipeline::Error, whose kind() (ErrorKind: SourceNotFound / SourceTooLarge / SourceUnreadable / Upstream / UpstreamTimeout / Undecodable / Internal) is the stable classification the server's own status mapping is built on — match on it instead of parsing messages, with a wildcard arm for kinds added later. Params also carries per-call overrides (webp_quality, png_effort, png_quantize, auto_rotate, icc, flatten_bg, linear_light, avif_quality, …) for the knobs that are otherwise process-global environment variables: None keeps the env-configured behavior, Some wins per call — so one process can run different settings side by side. Rustdoc examples on probe, process, and Params are compiled and run in CI; see also examples/:

cargo run --release --example thumbnail -- photo.jpg 300 200 out.jpg
cargo run --release --example transcode -- photo.jpg 800 800 webp out.webp
cargo run --release --example probe     -- photo.webp

Configuration

Everything is environment variables, read once at startup. The shared rule is fail-closed: any variable that is set but unparseable or out of range refuses to boot with a message naming it — a typo'd limit never silently falls back to a default.

Server

Sources

Encoding

Pixel pipeline

Deployment

Per-platform guides live in docs/:

• Docker / docker-compose — tag pinning (latest rebuilds on every main push), read-only mounts, remote origins, graceful docker stop, building tuned images.
• Kubernetes — an example Deployment with probes, resource limits (the worker count follows the cgroup CPU quota), security context, rolling-update drain behavior, and restoring cross-pod request coalescing via ingress URI hashing.
• Cloud Run & serverless containers — the PORT contract, remote-origin mode, gs:// with the service identity, concurrency-vs-vCPU sizing, and why per-process request coalescing yields nothing on scaled-out shapes.

The short version for every platform: pin an image version, put a CDN in front (responses carry a 1-year Cache-Control), give the process whole CPUs, and allow ≥10s of shutdown grace so in-flight encodes drain.

Not yet implemented (out of PoC scope)

• Private S3 / S3-compatible sources (gs:// landed in 0.7.4; s3:// is tracked in #11 and fails at boot with a pointer rather than misbehaving)
• JXL output (the @jxl token is reserved and returns a clear error)
• Animated output (animated AVIF and WebP sources render their first frame, like other image proxies)
• Response caching

Roadmap

Rough order, subject to change (experimental PoC):

• s3:// sources — S3 and S3-compatible endpoints (R2, MinIO, B2) with static credentials first, the AWS credential chain after (#11).
• Per-image output format selection — choose quantized-PNG vs WebP per image rather than per deployment (#6).
• JXL output once a maintained encoder binding stabilizes.
• Response caching (keyed on the resolved URL + format).

Status

Experimental PoC — APIs and the HTTP interface will change without notice. The @oximg npm package is a name reservation.

License

Apache-2.0.

The compiled binary statically links third-party code (jpegli/libjxl — BSD-3-Clause, Highway — Apache-2.0, mozjpeg/libjpeg-turbo — IJG). Their license texts and required notices are bundled in THIRD-PARTY-LICENSES.md, generated with cargo about. Dependency licensing is gated in CI by cargo deny (deny.toml).