AI Canvas

What. AI Canvas is a node-based editor for building multi-step AI image workflows in the browser. Wire a prompt into a generator, the result into an ImageMagick filter, that into a compositor — then save the graph as a workflow and re-run it from a simple form that asks only for the inputs that matter.

Why. ComfyUI is powerful, but it is local and it wants Python. I wanted the same graph model web-first and API-backed: nothing to install, and adding a new capability means registering an endpoint rather than shipping a custom node.

How. Thirty node types span inputs, math and vector operations, text generation, image generation and image-to-image, background removal, and seven ImageMagick operations (grid, resize, crop, rotate, composite, watermark, filter). Prompts declare variables with a $(name:type) syntax and those become typed pins on the node — a workflow's input form is derived from the prompts themselves. The Creator Suite builds custom nodes for a given service without writing code, and any OpenAI-compatible endpoint can be registered from settings. Execution walks the dependency graph with batch fan-out, a configurable parallel-run limit, per-item progress, and retry of only the failed items.

Providers wired in: OpenAI (DALL·E 2/3, gpt-image-1, chat), Google Gemini and Imagen — including Gemini 2.5 Flash Image, the "Nano Banana" model — for image-to-image, Clipdrop for background removal, and DoubTech.ai for async art jobs and ImageMagick.

Stack: Next.js 15, React 19, TypeScript, Genkit + Google GenAI, Firebase (Auth, Firestore, App Hosting), Tailwind + shadcn/ui.

Use Case #1 - Embossed Challenge Coin Generation

A recent hobby of mine has been fabrication of 2.5D challenge coins. Getting there means taking a central coin design out of an image service like MidJourney and working it into a usable depth map — several passes across different models, combining depth maps, edge detection, and a clean base image to get a quality output. AI Canvas collapses that into a single prompt, or a different workflow that takes an existing image as its starting point.

Built Sep 2025 – Mar 2026.