mrslimslim/gpt-image-canvas

What it solves

GPT Image Canvas provides a local-first creative workspace for AI image generation. It solves the problem of managing multiple AI-generated images and complex generation workflows by combining an infinite canvas (tldraw) with prompt management and agentic planning.

How it works

The system integrates a web-based canvas where users can arrange images and prompts. It uses a Hono API and SQLite for local storage of project state and history. For image generation, it connects to OpenAI-compatible endpoints (like gpt-image-2) or via a Codex login.

It features two primary workflows:

• Manual: Direct text-to-image or reference-image generation.
• Agent: An LLM-driven planner that takes a task description and optional reference images to create a DAG-based (Directed Acyclic Graph) execution plan for multi-image generation, which the system then executes in parallel or sequence.

Who it’s for

Digital artists, designers, and AI enthusiasts who want a visual, non-linear workspace to organize their AI image generation process and manage complex, multi-step image projects.

Highlights

• Infinite Canvas: Uses tldraw for a spatial layout of generated assets.
• Agentic Planning: Multi-image tasks are broken down into a DAG of generation jobs.
• Reference-Image Support: Ability to use existing canvas images as references for new generations.
• Local-First: Project state, history, and assets are stored locally in SQLite and a local folder.
• Flexible Providers: Supports OpenAI, OpenAI-compatible APIs, and Codex login.
• Cloud Backup: Optional integration with S3-compatible storage, Cloudflare R2, or Tencent Cloud COS.