Web app · AI images · Open source

Cut the recipe once. Run it again and again.

Trafaret is a node-based editor for AI images. Put a prompt, a photo, Generate, Edit and Result blocks on a canvas, wire them up, pick a style and press Run. Independent branches render side by side, a failed block retries on its own, and the next run redraws only what changed. Trafaret is Russian for stencil.

Open the live demo → Source on GitHub

The demo runs on a free instance: the first visit after a quiet spell takes about a minute to wake it up.

The Trafaret editor: a prompt wired into two Seedream 4.5 Generate blocks with their results, a photo edited into a new scene, and a run timeline at the bottom — A prompt rendered by two Seedream 4.5 branches, plus a photo edit. On this run only the edit was redrawn — the timeline reads 1 AI call · 2 reused · saved ~34.5 s.

Build once, re-run only what changed

Results are kept by what made them

Every block's result is stored under a fingerprint of its inputs: the prompt, the model, the style, the seed and everything upstream. Change one style and only its branch goes back to the AI; the rest comes back instantly, and the timeline says how much time that saved. Want a new take? 🎲 gives a block a new seed.

Branches run side by side, failures stay small

A block starts the moment its own inputs are ready, so independent branches overlap instead of waiting in line. When one AI call fails, only the blocks that need it stop; Retry runs that part again without touching the rest. Nothing paid is ever retried behind your back.

Styles live on the server

A style preset is a prompt, things to avoid and reference images. The server merges it into each request at run time, and the inspector shows exactly what was sent to the AI — the final prompt and every image, with its role.

What the templates make

Real FLUX.2 klein output from the demo's templates: one prompt in two styles, and a casual product photo moved into a new scene.

A white mug with rising steam on a wooden table, clean and minimal, pastel backdrop — Premium 3D“A ceramic coffee mug with steam rising, on a wooden table”

The same kind of mug on a sunlit kitchen counter next to a linen towel — LifestyleThe same prompt, the other style

A sage-green ceramic teapot on a cluttered dining table with mail, keys and a laptop — PhotoThe sample product shot

The same teapot on a sunlit marble kitchen counter — Edit“Put the product on a sunlit marble kitchen counter”

Under the hood

EditorReact 19 and React Flow, organised by Feature-Sliced Design. The canvas checks every wire as you drag it, and the workflow saves itself in your browser.
ServerFastify in TypeScript. It validates the graph with the same rules as the editor, runs it block by block as inputs become ready, and caches each result by its inputs.
ModelsFLUX.2 klein on Cloudflare Workers AI, Seedream 4.5 on BytePlus ModelArk, and a Mock model that needs no key — within small daily caps on the demo.
DeliveryOne Docker image on Render. Every change goes through unit tests, browser tests and an automated accessibility scan before it deploys.

A workflow is plain JSON: export it from the editor and post it to /api/runs with curl. The API reference is live too.

No accounts, no tracking

There is nothing to sign up for, and no cookies or analytics. A prompt or photo you run goes to the provider of the model you picked; the server keeps it in memory only, until it restarts. Your workflow stays in your own browser. The privacy policy says exactly that.