ima2-gen runs image generation on your own machine

Most people who generate images or video with AI today do it inside one vendor's web interface. ima2-gen, a project by the developer lidge-jun, inverts that arrangement: it installs a runtime and a studio on the user's own machine and reaches out to several providers from there. The workspace opens in a browser at localhost:3333, and the code carries an MIT license.

What the tool is meant to do

On the image side the project lists OpenAI, Grok, Gemini, NovelAI and AtlasCloud, connected through either OAuth or an API key. Video generation runs mainly through Grok, with Runway and Higgsfield attached over MCP. A storyboard mode is supposed to hold consecutive frames together rather than treating each one as an unrelated render.

There is also a canvas mode for annotating, erasing and knocking out backgrounds in a single step, plus a local gallery that records history per session. A single job accepts up to five image references or seven video references. The workflows themselves come in three shapes: classic, node-based and multimode.

Built for coding agents too

The second audience is the more unusual one. Alongside human users, the project addresses coding agents directly and bundles three ready-made skills for them, named Core, Frontend and UI/UX Design and invoked through their own subcommands. They are described as instructions covering the generation protocol, asset pipelines and design workflows.

An MCP Registry hook lets external tools plug in. That places ima2-gen among a growing set of tools that offer not only a human interface but a machine-readable one, on the assumption that an automated development environment will be a regular caller rather than an exception.

Installation and stated limits

Three installation routes are documented: a global npm install, Docker with a compose file, or platform scripts for macOS, Windows and Linux or WSL that are meant to work without npm at all. The project is written in TypeScript.

The documentation is unusually direct about its own gaps. NovelAI is text-to-image only, with no references, edits or masks. The Antigravity path emits a fixed 1024 × 1024 JPEG and handles at most three references. HEIC and HEIF files are not accepted through the browser path.

What has not been verified

Everything above comes from the project's own description. No second independent newsroom has confirmed it, so whether the listed provider paths work as advertised in daily use remains unchecked here.

The date of the most recent commit and any specific release version were also not established. The figures of 722 stars and 120 forks are a snapshot from the moment of retrieval and can shift from day to day.