zenjpeg

High-quality JPEG encoder combining the best techniques from mozjpeg and jpegli for Pareto-optimal compression.

Overview

zenjpeg automatically selects the optimal encoding strategy based on quality requirements:

• Low quality (Q < 70): Uses mozjpeg-style trellis quantization for best compression
• High quality (Q >= 70): Uses jpegli-style adaptive quantization for perceptual quality
• Quality targeting: Binary search to hit specific SSIMULACRA2/Butteraugli targets

Features

• Adaptive strategy selection - Automatically picks best encoder based on target quality
• Trellis quantization - Rate-distortion optimized coefficient selection (from mozjpeg)
• Adaptive quantization - Content-aware bit allocation (from jpegli)
• Progressive encoding - Multiple scan scripts for optimized progressive display
• Huffman optimization - Two-pass encoding for optimal entropy coding
• Perceptual targeting - Target specific quality metrics (SSIMULACRA2, Butteraugli, DSSIM)

Usage

use zenjpeg::{Encoder, Quality};

// Basic encoding
let encoder = Encoder::new().quality(Quality::Standard(85));
let jpeg_data = encoder.encode_rgb(&pixels, width, height)?;

// Maximum compression (progressive + trellis)
let encoder = Encoder::max_compression();
let jpeg_data = encoder.encode_rgb(&pixels, width, height)?;

// Maximum quality
let encoder = Encoder::max_quality();
let jpeg_data = encoder.encode_rgb(&pixels, width, height)?;

Feature Flags

• simd - Enable SIMD intrinsics for faster encoding
• cms-lcms2 - Use Little CMS 2 for ICC color management
• cms-moxcms - Use moxcms (pure Rust) for ICC color management

Performance

At SSIM2 >= 80 quality target:

• jpegli: 1.310 bpp (best efficiency)
• mozjpeg-oxide: 1.437 bpp
• zenjpeg: 1.458 bpp (within 1.5% of mozjpeg-oxide)

zenjpeg appears on the Pareto front at multiple quality levels, offering the best of both encoders depending on the target.

Requirements

• Rust 1.70.0 or later

License

AGPL-3.0-or-later

AI-Generated Code Notice

This crate was developed with assistance from Claude (Anthropic). While thoroughly tested against reference implementations (mozjpeg, jpegli) with 200+ passing tests, not all code has been manually reviewed.

Before production use:

• Review critical paths for your use case
• Run your own validation on representative images
• Consider the AGPL license implications