Chatgpt Image

ChatGPT image generation (board-first)

Project home: ${CHATGPT_IMAGE_GEN_HOME:-~/Workspace/chatgpt-endless-canvas} (referred to as $PROJ below — resolve it once per session). Data root: ~/Documents/chatgpt-endless-image-gen/.

Bootstrap (skill install ≠ project install). This file only teaches the workflow; the project code + venv must exist at $PROJ. Check once per session, and set it up if missing:

[ -f "$PROJ/board_server.py" ] || {
  git clone https://github.com/ManiacMike/chatgpt-endless-canvas.git "$PROJ" &&
  python3 -m venv "$PROJ/.venv" &&
  "$PROJ/.venv/bin/pip" install -r "$PROJ/requirements.txt"
}

Core rule: submit ALL generation requests through the board server API (POST /api/generate) instead of invoking the Python script directly — the server serializes jobs, spaces batches with random pauses (rate-limit protection), streams results onto the canvas, and records lineage. Call the script directly only as a fallback when the server cannot run, or for --grab-only recovery.

Standard flow (every request)

1. Debug Chrome ready?

   curl -s --max-time 2 http://127.0.0.1:9222/json/version

   On failure → ask the user to run bash $PROJ/launch-chrome-debug.sh, log into chatgpt.com in the opened window, and keep it open. Do NOT run it for them in the background — the window needs to stay interactive.

2. Board server running? Use the idempotent launcher (safe from any session; already-running → no-op; starts detached via nohup so it outlives this session):

   bash $PROJ/start.sh

   It prints the ACTUAL url (board started: http://127.0.0.1:<port>) — the port may NOT be 8090: ports occupied by other programs (including stale pre-rename copies of this project) are skipped automatically. Use the printed url as $BOARD for ALL API calls below, and open $BOARD if the user doesn't have the board open. To re-discover it later, GET /api/health must return "app": "chatgpt-endless-canvas" (also in ~/Documents/chatgpt-endless-image-gen/server.json alongside pid/port) — a port that answers /api/state but not /api/health is NOT this server. Do NOT run board_server.py with run_in_background — that ties the server to this session. Log: ~/Documents/chatgpt-endless-image-gen/board.log. Unfinished jobs survive a server restart (persisted + requeued), so if the environment kills the detached process, just rerun start.sh.

3. Submit (POST /api/generate, pick by scenario):

   • Plain: {"prompt": "specific English description"}
   • Multiple images: {"prompts": ["...", ...]} (≤50; the server queues them with random 30–120 s gaps and runs up to 3 in parallel (BOARD_WORKERS), each in its own dedicated ChatGPT tab — ALWAYS use this for batches, never loop yourself. The debug Chrome holding several chatgpt.com tabs is expected; don't close them.)
   • Style reference from a board image: add "name": "<filename>"
   • Local file as reference: first POST /api/upload?name=<file> with raw bytes (curl --data-binary @file.png), then generate with the returned name
   • Edit by annotations: POST /api/regenerate {"name":"<image>"} — needs pending annotations (you may write them for the user via POST /api/annotations: {image: [{id,x,y,w,h,note,status:"pending"}]}, coords normalized 0-1, point marks have w=h=0)
   • Delete from the board: POST /api/delete {"name":"<image>"} — removes the file and its layout/annotation/lineage entries (409 while a job is still using the image; confirm with the user before deleting)

4. Await results: poll GET /api/state → jobs until the job is done (output = filename) or error (reason included). ~1-5 min per image. For batches don't block — tell the user images will appear on the board as they finish. Files land in the active board dir (state.board.dir). Each job records conversationId — the ChatGPT chat uuid it ran in (its stable identity; interrupted jobs are auto-recovered from that conversation on server restart instead of regenerating).