BlazeDiff is a high-performance diff ecosystem. Originally built in JavaScript as a pixel-perfect image comparison library that's 1.5x faster than pixelmatch. Now, BlazeDiff has evolved into a comprehensive suite of blazing-fast diff tools including image comparison, image diff analysis deterministically + agent-in-the-loop verdict, object diffing, perceptual quality metrics, framework-agnostic UI renderers, and React components for visualizing differences.

Available Packages

Core Libraries

@blazediff/core-native - The fastest single-threaded image diff in the world (Rust + SIMD, 4.4-4.9x faster than odiff on 4K images)
@blazediff/core-wasm - WebAssembly build of the same Rust algorithm (wasm32 + v128 SIMD, ~51% faster than pixelmatch). For browsers, edge runtimes, and any wasm host.
@blazediff/core - Pixel-perfect image comparison (1.5x faster than pixelmatch)
@blazediff/interpret-native - Native Rust structured interpretation of a diff: what changed in each region, where, and how much - not just which pixels differ
@blazediff/object - High-performance object diffing with detailed change tracking
@blazediff/ssim - SSIM, MS-SSIM, and Hitchhiker's SSIM for perceptual quality assessment
@blazediff/ssim-native - Native Rust build of the same metrics plus perceptual SSIM (~15x faster than the JS port; ~4x for Hitchhiker's)
@blazediff/gmsd - Gradient Magnitude Similarity Deviation metric
@blazediff/milo-native - Native Rust MILO (Çoğalan et al. 2025), a learned perceptual metric that models visual masking. The authors' weights embedded, PyTorch parity, no PyTorch
@blazediff/milo-wasm - WebAssembly build of the same metric for browsers, edge runtimes, and any wasm host
@blazediff/codec-png-native - The blazediff-png Rust PNG codec for Node.js through N-API: spng-parity decode, libdeflate encode
@blazediff/codec-png-wasm - WebAssembly build of the same codec for browsers, edge runtimes, and any wasm host

Rust Crates

blazediff - The Rust crate powering @blazediff/core-native: block-based image diffing with SIMD (source)
blazediff-png - From-scratch PNG codec in Rust, single-threaded and SIMD-first, with byte-exact decode parity to libspng and faster on every fixture. Powers @blazediff/codec-png-native and @blazediff/codec-png-wasm (source)
blazediff-ssim - SSIM, MS-SSIM and Hitchhiker's SSIM in Rust, dependency-free and SIMD-vectorised, held to the reference MATLAB scripts through Octave (source)
blazediff-interpret - Structured region analysis for image diffs: classify what changed, not just where. Powers @blazediff/interpret-native (source)
blazediff-milo - MILO perceptual quality metric as a std-only SIMD CNN, multi-threaded, held to the authors' PyTorch outputs by test. Powers @blazediff/milo-native and @blazediff/milo-wasm (source)

Python Packages

blazediff - Python bindings (via maturin) for the Rust diff core (source)
blazediff-ssim - Python bindings for the SSIM family, imported as blazediff_ssim (source)
blazediff-interpret - Python bindings for the region classifier, imported as blazediff_interpret (source)

Command Line Tools

@blazediff/cli - JS CLI with multiple algorithms (diff, GMSD, SSIM, MS-SSIM, Hitchhiker's SSIM)
@blazediff/agent - Agentic visual regression. Auto-discovers routes, captures deterministic screenshots, and hands ambiguous diffs back to your coding agent (Claude Code, Cursor, Codex) to judge.

UI Components

@blazediff/react - React components for diff visualization
@blazediff/ui - Headless engine + framework-agnostic renderer for image-diff UIs

Quick Links

Documentation - Complete API reference and guides
Examples - Interactive demos and code samples
Benchmarks - Performance comparisons and metrics

Installation

BlazeDiff packages are dual-published. Install from whichever registry fits your runtime.

# Node / Bun via npm
npm install @blazediff/core
bun i @blazediff/core

# Deno / Bun via JSR
deno add jsr:@blazediff/core
bunx jsr add @blazediff/core

# Rust via Cargo
cargo add blazediff

# Python via pip
pip install blazediff blazediff-ssim blazediff-interpret

Every package above is available on both registries except the test-runner adapters (@blazediff/vitest, @blazediff/jest, @blazediff/bun) and the UI libraries (@blazediff/ui, @blazediff/react), which remain NPM-only - the adapters augment each runner's Matchers types and the UI packages have browser-DOM and React type surfaces that JSR's publish-time slow-types check doesn't allow. Native-binary sub-packages under @blazediff/core-native-*, @blazediff/ssim-native-*, @blazediff/interpret-native-*, @blazediff/milo-native-* and @blazediff/codec-png-native-* are also NPM-only; Deno consumers resolve them transparently via npm: specifiers declared inside the parent package.

Performance

BlazeDiff delivers significant performance improvements across all components:

Native (Rust): 4.4-4.9x faster than odiff on 4K images (5.7-6.7x from encoded buffers)
WASM: ~51% faster than pixelmatch, up to ~10x on 4K (browser, edge, any wasm host)
Image Pixel-by-Pixel (JS): ~50% faster than pixelmatch (up to 88% on identical images)
SSIM (JS): ~30% faster than ssim.js, ~75% faster with Hitchhiker's SSIM
SSIM (Rust): ~15x faster again than the JS port (~4x for Hitchhiker's SSIM), from the same decoded buffers
Object Diff: ~55% faster than microdiff (up to 96% on identical arrays)

View Detailed Benchmarks - Complete performance data, test methodology, and hardware specifications.

Contributing

Contributions are welcome! Please see the Contributing Guide for details.

License

MIT License - see LICENSE file for details.

Algorithm Licenses

The @blazediff/ssim and @blazediff/gmsd packages implement perceptual quality metrics based on published research. See the licenses folder for detailed attribution and licensing information:

SSIM - Zhou Wang et al., IEEE 2004
MS-SSIM - Zhou Wang et al., Asilomar 2003
Hitchhiker's SSIM - Venkataramanan et al., IEEE Access 2021

Acknowledgments

Image diffing built on the excellent pixelmatch algorithm
SSIM and MS-SSIM based on groundbreaking research by Zhou Wang, Alan C. Bovik, and colleagues
Hitchhiker's SSIM based on research by Venkataramanan, Wu, Bovik, Katsavounidis, and Shahid
GMSD based on research by Xue, Zhang, Mou, and Bovik

Built for high-performance difference detection across images and data structures