# Magenta.js: browser-based inference on Magenta's music, sketch, and image models > Magenta.js is a set of 3 TypeScript npm packages for running pre-trained Magenta models in a browser via TensorFlow.js, covering music note generation, sketch drawing, and image style transfer. All 3 packages handle inference only; training and custom model weights require the Python Magenta repository instead. **magenta/magenta-js** — Magenta.js: Music and Art Generation with Machine Learning in the browser - Repository: https://github.com/magenta/magenta-js - Website: https://magenta.tensorflow.org - Stars: 2,133 · Forks: 335 - Language: TypeScript - License: Apache-2.0 - Published: 2026-10-11 · Updated: 2026-10-11 · Language: en - Canonical page: https://hysenlabs.com/projects/magenta-magenta-js ## Inference with pre-trained Magenta models in the browser, not model training Magenta.js sits at the inference end of the Magenta ecosystem. Running a generative model in Python typically requires a local environment, a checkpoint file, and often a GPU. Magenta.js moves the inference step to the browser using TensorFlow.js, letting the same generative models run on the client without leaving JavaScript. Setup is adding an npm package. What the library does not cover is training. All 3 packages operate on pre-trained weights from the broader Magenta project. Writing your own training loop, fine-tuning existing models, or providing custom checkpoint files is outside the scope of Magenta.js. A project that needs those capabilities still requires the Python Magenta repository at github.com/tensorflow/magenta. ## Five named models in @magenta/music: MusicVAE, MelodyRNN, DrumsRNN, PerformanceRNN, and ImprovRNN The @magenta/music package holds TensorFlow.js implementations of 5 note-based music generation models. MusicVAE, MelodyRNN, DrumsRNN, PerformanceRNN, and ImprovRNN are each named in the README, identifying the distinct generative tasks the package handles. What each model produces, how they differ from one another, and which one to choose for a given task are not explained in the README. Complete API documentation for all 5 is at magenta.github.io/magenta-js, which is where usage specifics live. Selecting between MelodyRNN and MusicVAE for a melody continuation task requires reading those docs rather than the README. The package description specifies note-based models. Whether audio synthesis, raw waveforms, or MIDI file handling are included in the package is not addressed in the README; the description establishes an abstraction level (notes) without specifying the full audio pipeline. ## SketchRNN and Arbitrary Style Transfer in @magenta/sketch and @magenta/image Beyond music, 2 additional packages handle visual generation. @magenta/sketch contains TensorFlow.js support for SketchRNN, a model that generates human-style line drawings. @magenta/image covers Arbitrary Style Transfer, a model that applies the visual style of one image to the content of another. Both packages follow the same pattern as @magenta/music: TypeScript implementations built on TensorFlow.js, published as npm packages, and requiring pre-trained weights that the README does not describe how to obtain. Each occupies its own subdirectory at the top level of the repository: sketch/ and image/ alongside music/. Example applications showing all 3 packages in action are available at g.co/magenta/js, which the README identifies as the place to find working demonstrations. ## Installing the @magenta npm packages and finding the API reference Each package in the @magenta scope is published independently on npm. A project that needs only music generation can add @magenta/music without pulling in @magenta/sketch or @magenta/image. All 3 are findable under the @magenta scope at npmjs.com, and the README links the full listing at https://www.npmjs.com/search?q=%40magenta. Installing all 3 at once adds more JavaScript than most projects need, since each package carries its own TensorFlow.js model logic. Complete API documentation for all packages is at https://magenta.github.io/magenta-js. Example applications with working code are at g.co/magenta/js. Adding a package alone does not give you a checkpoint URL or a loading example; those come from the external documentation, not from the package itself. ## What the README omits: model weights and how to load them Five named models appear in the @magenta/music README section (MusicVAE, MelodyRNN, DrumsRNN, PerformanceRNN, ImprovRNN), but the README says nothing about where their pre-trained weights come from or how they are loaded into a running instance. Inference on a generative model requires both the architecture, which the npm package provides, and the weights, which the README treats as already available without linking to them. A developer who reads the README and installs the package will be stuck at the weight-loading step with no guidance from within the README itself. The documentation at magenta.github.io/magenta-js and the examples at g.co/magenta/js are the correct resources for checkpoint URLs and loading code. The README explicitly delegates all content past the package list to those 2 external sources. ## Magenta.js versus the Python Magenta project Magenta.js and the Python Magenta project at github.com/tensorflow/magenta cover different parts of the ML workflow. Python Magenta handles training, fine-tuning, and producing the checkpoint files that Magenta.js then consumes. Magenta.js handles running those checkpoints in a browser without requiring Python, Node, or GPU hardware on the client. A project that wants to train a custom melody model or adjust an existing one still starts in Python. Magenta.js is the deployment end: it moves a finished model to web users through a browser page, where installing a Python runtime is not an option. That split also explains why Magenta.js is scoped to pre-trained models. Training in JavaScript is possible with TensorFlow.js directly, but Magenta.js does not wrap that capability, and neither the README nor the library description presents it as a training tool. ## When Magenta.js does not fit: runtime, audio pipeline, and versioning gaps Magenta.js requires JavaScript and either a browser or a Node environment. Projects targeting Python pipelines, embedded audio hardware, or platforms without a JavaScript runtime get nothing from this library. Generating audio waveforms directly is not what the packages cover. @magenta/music describes itself as built around note-based models, which is a specific musical abstraction. A project that needs raw audio synthesis faces a gap the README does not address. Carrying an Apache-2.0 designation, the project had its last push on 2026-06-22, with no GitHub releases published. A developer who wants to pin to a specific source state has no tagged version available; npm package versions are the only versioning surface, and pinning in package.json is the recommended approach when stability matters. ## Conclusion Magenta.js is worth adopting when you want browser-based music generation or image style transfer without a Python backend. Avoid it if you need to train models or provide your own weights: the library is inference-only, and the README says nothing about checkpoint loading. Before building on it, read the documentation at magenta.github.io/magenta-js to understand which models are available and what checkpoint format each expects, since that detail is absent from the README and the npm packages alone. ## FAQ ### What is Magenta.js? Magenta.js is a collection of TypeScript npm packages for running pre-trained Magenta generative models in the browser using TensorFlow.js. It covers music note generation, sketch drawing, and image style transfer, and requires no Python backend. ### How do I install Magenta.js? Each package is installed separately via npm: @magenta/music for music models, @magenta/sketch for SketchRNN, and @magenta/image for Arbitrary Style Transfer. A complete tutorial and checkpoint loading instructions are at magenta.github.io/magenta-js. ### What music models does @magenta/music include? The README names 5 models in @magenta/music: MusicVAE, MelodyRNN, DrumsRNN, PerformanceRNN, and ImprovRNN. All are note-based models run via TensorFlow.js in the browser. ### Can I use Magenta.js without Python? Yes. Magenta.js runs in the browser with no Python backend required. It handles only inference on pre-trained models; training or fine-tuning models still requires the Python Magenta project at github.com/tensorflow/magenta. ### What is the difference between Magenta.js and the Python Magenta project? Python Magenta handles model training and produces checkpoint files. Magenta.js consumes those checkpoints to run inference in the browser. Magenta.js does not expose any training capability. ## Sources - [Issues](https://github.com/magenta/magenta-js/issues) - [License: Apache-2.0](https://github.com/magenta/magenta-js/blob/master/LICENSE) - [magenta/magenta-js on GitHub](https://github.com/magenta/magenta-js) - [Project website](https://magenta.tensorflow.org) - [README](https://github.com/magenta/magenta-js/blob/master/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/magenta-magenta-js