rusty_jpeg

Pure-Rust JPEG / MJPEG decoder + encoder. No C, no FFI. Baseline and progressive DCT, planar YUV in and out, with real quality and chroma-subsampling control.

This crate is a vendored merge of two upstream pure-Rust projects, carried forward in-tree as one codec:

See NOTICE.md for attribution and the IJG obligations, and CHANGES.md for everything changed since vendoring.

This software is based in part on the work of the Independent JPEG Group.

Performance vs FFmpeg

Both halves are measured against system FFmpeg 8.1.2 on the same machine, one pinned core each, CPU time (not wall), best-of-N with the arms alternated. Content is a 1920×1080 4:2:0 photographic fixture (1/f fractal noise, q90, 11.8× compression) — synthetic clips misprice JPEG stage shares badly, so this one is calibrated to real photographic coefficient density.

Both numbers are deliberately unflattering to us where the methodology allows a choice:

• Encode is compared at matched output size, not matched -q:v. At equal -q:v our files are 41% larger, which is a different operating point and would price our extra bits as speed. Ours at -q:v 5 is 27.17 MB against FFmpeg's 28.04 MB at -q:v 3 — we are slightly smaller and still faster. Both run fixed Huffman tables in one pass, because our CLI defaults to optimized tables (two passes) and FFmpeg's does not; comparing defaults would price our better compression as a speed loss.
• Decode is compared on a byte-identical bitstream — the reference clip is stream-copied from the exact fixture and cmp-gated before timing, after an earlier harness was caught feeding FFmpeg a 2.67× smaller file. It is a paired win-rate with a z-score rather than a ratio of medians, because on this box the medians and the minima of the same arms disagreed by six points. The honest read is parity; we do not claim the +2.8% the median suggests.

Where the speed comes from, all gated on byte-identical output:

• an AVX2 forward-DCT + quantize kernel in the encoder;
• an AVX2 IDCT that transforms two 8×8 blocks per instruction stream (block A in the low 128-bit lane, block B in the high one — every op involved is lane-independent, so it is byte-identical to running the SSSE3 kernel twice, which a 64-round oracle test asserts). Worth 7.8% of whole decode;
• a whole-block DC-only shortcut ahead of the SIMD dispatch (~31% of blocks on photographic content have no AC energy);
• buffered entropy reads with a bulk 8-byte refill, and recycled output planes and MCU-row coefficient buffers;
• no rayon. Upstream jpeg-decoder enables it by default; measured here it is a net loss at every image size — turning it off took 1080p decode from 6.94 to 3.82 ms/frame.

Decoder::set_single_threaded(true) selects the synchronous worker, which uses ~38% less CPU than the threaded one; the threaded default is still the faster choice in wall-clock on a multi-core box (6.97 vs 7.96 ms/frame).

Why merge them

Holding the two halves in one crate buys three things the separate crates could not:

1. The encoder is gated against the decoder as a round-trip oracle — the standing correctness gate for every change, running in CI on real content rather than as a one-off bring-up check.
2. Shared primitives — quantization tables, zig-zag order, and the feature-gated stage profiler are defined once.
3. The encoder's SIMD is on by default. Upstream put its AVX2 FDCT and RGB→YCbCr kernels behind a non-default simd feature; consumers who took the default (as this workspace did) silently ran the scalar path.

Decode

use std::io::Cursor;

fn main() -> Result<(), rusty_jpeg::decode::Error> {
    let bytes = std::fs::read("in.jpg").expect("read input");

    let mut decoder = rusty_jpeg::Decoder::new(Cursor::new(bytes));
    let pixels = decoder.decode()?;
    let info = decoder.info().expect("populated by decode()");

    println!("{}x{}, {:?}", info.width, info.height, info.pixel_format);
    println!("{} bytes of pixel data", pixels.len());
    Ok(())
}

Encode

use rusty_jpeg::{ColorType, Encoder};

fn main() -> Result<(), rusty_jpeg::EncodingError> {
    // A 2x2 RGB image: red, green, blue, white.
    let pixels = [255, 0, 0, 0, 255, 0, 0, 0, 255, 255, 255, 255];

    let mut jpeg = Vec::new();
    let encoder = Encoder::new(&mut jpeg, 90); // quality 0-100
    encoder.encode(&pixels, 2, 2, ColorType::Rgb)?;

    std::fs::write("out.jpg", jpeg).expect("write output");
    Ok(())
}

Chroma subsampling (SamplingFactor) and quantization table selection (QuantizationTableType) are set on the Encoder before encode.

Features

Part of Remade With Rust

This crate is the standalone JPEG engine of remade_ffmpeg_rs — a ground-up, permissively-licensed Rust rebuild of FFmpeg: a drop-in ffmpeg/ffprobe CLI on pure-Rust codecs, with no copyleft. Also check out our sister project FFAI — media for an AI-first world — and the rest of github.com/remade-with-rust, including the sibling codec crates rusty_h264, rusty_vp9, rusty_mp3, rusty_aac, rusty-opus, rusty_vorbis, and the rusty-av1-toolkit forks.

About Mata Network

Mata Network builds sovereign, self-hostable infrastructure. Remade With Rust is our open-source home for the permissively-licensed building blocks that work depends on.

Licence

(MIT OR Apache-2.0) AND IJG — the IJG clause attaches to the forward-DCT files inherited from the encoder. See NOTICE.md.