comic-maker — one sentence → a lettered comic, fully local

Type a story idea. A local LLM writes the script, a local image model draws every panel with a consistent character in a consistent style, and a compositor letters the pages — speech bubbles, SFX, diagonal panel cuts, title strip, CBZ export. No cloud, no credits, no accounts. Built and used daily by the Aillex project, where an AI runs her own comic studio.

⚠️ Actively developing. This is an early cut published as we build it in the open — expect the API, prompts, and layouts to keep improving. Watch the repo (or the channel) to follow along; issues and ideas welcome.

How it works

"a small robot learns to make coffee"
   │  script    local LLM (Ollama) → title, characters w/ locked visual "looks",
   │            pages & panels: scene / emotion / dialogue / SFX   (edit the JSON if you like)
   │  panels    ComfyUI (z-image turbo) → one image per panel
   │            consistency = same character → same seed family + full look restated per panel
   │  assemble  PIL compositor → dynamic layouts (1–9 panels/page), diagonal gutters,
   │            wrapped bubbles w/ tails, Impact SFX, page-1 title, .cbz
   ▼
page_01.png … + your_story.cbz

Quickstart

Prereqs: Python 3.10+ with pillow, Ollama with any chat model, ComfyUI with the z-image turbo checkpoints (or edit comfy_gen for your favorite model — it's ~30 lines).

python comicify.py new "a lighthouse keeper discovers the fog is singing" --style noir --pages 2
python comicify.py all a_lighthouse_keeper_discovers_th

Stages run independently too (script → edit script.json by hand → panels → assemble).

Config via env: COMIC_ROOT (project/output dirs), OLLAMA_URL, COMFY_URL, COMIC_FONTS.

Styles

sunday-strip (default) · manga · noir · euro-album — or pass any style description string. Tip we learned the hard way: never put nameable words in a style ("sunday", "newspaper") — the model will happily paint them into your art.

Consistent characters

Define a cast entry (see CAST in comicify.py) with a full, repeatable visual description; the script stage forces the LLM to star it and every panel prompt restates the look with a per-character seed family. That combination — not LoRA training — is what keeps the character recognizable panel to panel. (For a trained character that survives any style, see our character LoRA guide.)

Examples

Roadmap (the "will develop further" part)

• speech-bubble placement aware of character positions (bubbles currently anchor top-center)
• multi-character scenes with per-character bubble tails
• style LoRA slots for trained characters
• non-rectangular pages: splash panels, insets, bleeds
• agent mode: our Discord assistant already builds comics on request — packaging that pattern

MIT. Made with local models on one PC. — askaillex.com