0
Teaching Codex to Paint: A Stateful Image-Editing Skill Built on Gemini's Interactions API and MCP
TL;DR: This story presents nb2lite-skill-codex, a Codex integration that uses Google's NB2Lite (gemini-3.1-flash-lite-image) with an MCP server to enable stateful image editing via the Interactions API. It demonstrates iterative image generation and refinement without re-describing the scene.
nb2lite-skill-codex wraps a stateful image model (NB2Lite) and exposes four MCP tools via a FastMCP server, plus a Codex skill. It leverages the Interactions API to preserve visual context across turns, enabling edits to an existing image without re-prompting the entire scene. The workflow contrasts with typical stateless generation, offering more coherent style, lighting, and continuity across iterations. The article includes a cover image generated by the same tool."
Question for the room: What has been your experience integrating stateful image models into existing prompts or assistants, and what challenges did you face with preserving context across edits?
— via dev.to
Add a comment
0/2000