basketikun/infinite-canvas

The problem

Iterating on AI images is usually a linear mess: prompt in a chat box, download the output, drag it into a folder, re-upload it as a reference for the next edit. As soon as you have ten variations and three attached reference images, you lose track of which prompt produced which result. infinite-canvas puts that loop on an infinite spatial canvas where generated images, reference images, conversations and upstream results sit next to each other as nodes.

What's different

Most people solve this with a chat client plus a folder of downloads, or with node-graph tools like ComfyUI. This project sits in between: it is a canvas editor first, and a model frontend second. Three things actually stand out. It calls OpenAI-compatible endpoints directly from the browser, so your keys and history never touch a server you run. It treats edits as canvas operations on selected or upstream nodes, with results inserted back into the graph rather than replacing a chat turn. And it has a local Agent bridge (MCP) so Codex or Claude Code can manipulate the current canvas, plus a plugin system with a TypeScript SDK for custom node types. The video generation and audio features listed in the README are described only briefly; treat them as advertised rather than proven.

Who it's for

Solo creators, designers and prompt tinkerers who generate a lot of images and want a persistent visual workspace they host themselves. It fits people who already have an OpenAI-compatible relay or self-hosted API and are comfortable configuring base URLs and custom request scripts. Look elsewhere if you need stable data formats for a team, multi-user collaboration, or a product you can build on commercially without surprises — the README itself warns that local storage formats may change without backward compatibility. If your workflow is strictly reproducible pipelines rather than visual exploration, ComfyUI's graph-and-model approach is a better fit.

Cost of adoption

You need a Node/Bun toolchain or Docker; the quick start uses bun install and bun run dev under web/, and docker compose up -d serving port 3000. No GPU is required because inference is delegated to your API provider — so the real recurring cost is per-image API spend, and several of the README's sponsors are exactly those relay services. The bigger risk is maturity: the project is in development and explicitly disclaims data compatibility. State lives in the browser (IndexedDB and local storage), so backing up or migrating is on you. MIT license removes legal friction.

Verdict

This is a young project: created in May 2026, roughly five months old at the time of tracking, with 8,996 stars and a single trending-board appearance at +549 stars in one day. That pattern — a large star count, a 2,108-fork ratio, and one spike day — reads more like a community-driven burst than a long-maintained tool. The feature set is real and the browser-direct-to-API design is genuinely convenient, but the README's own stability warning should carry weight. Worth trying if you want a self-hosted canvas for image iteration and can tolerate churn; not yet something to standardize a team on.