Flux's copy-pasteable gallery of working neural nets

What it does

The repository collects runnable demonstrations of the Flux machine learning library, sorted into vision, text, and contributed miscellany. Each example is a self-contained Julia project pinned to specific package versions, so resurrecting an old diffusion model or VAE is meant to be frictionless. It reads as a reference shelf rather than a framework.

The interesting bit

The maintainers treat this as living documentation: models must include READMEs explaining architecture and results, should be CPU/GPU agnostic, and are explicitly graded on whether they use current Flux APIs. Some entries still carry compatibility warnings for ancient Flux versions, which doubles as an accidental museum of Julia ML history.

Key highlights

• Self-contained Julia projects with locked Manifests, so dependency drift is theoretically contained
• Coverage spans MLPs and LeNet up through diffusion models, GANs, VAEs, and a NanoGPT implementation
• GPU support is baked into most vision models via gpu calls
• Includes a “60 Minute Blitz” tutorial and transfer-learning examples
• Contribution guidelines demand short, clean code with minimal boilerplate

Caveats

• Version rot is visible: several models are stuck at Flux v0.11 or older, and the README itself notes that bringing examples current is an ongoing chore
• Gitpod support is listed but explicitly disclaimed as unofficial

Verdict

Worth browsing if you are learning Flux or need a verified starting point for a standard architecture. Skip it if you want a unified framework; this is explicitly a collection of loose scripts, not a product.