ComicCraft AI · An editable comic creation workspace

Try editing speech bubbles in your browser · If the demo is useful, star ComicCraft AI on GitHub.

Play the promo online · Watch the project demo (MP4) · Download the editable HyperFrames project

简体中文 · English

Turn a story into panels you can keep editing: scripts, character outfits, dialogue bubbles, masked repairs and chapter versions. Built with React, FastAPI, LangGraph, SQLite and Pillow.

Showcase · Windows installation · User guide · Architecture

Feature walkthrough

Screenshots from the running application. Open an image at full size, or try the interactive editor from the featured chapter.

Full showcase · Try the balloon editor

Creation	Planning and production
Novel import and adaptation Import a source, preview excerpts, and set the chapter’s panel count and required facts.	Visual references and storyboard draft Assign references per panel and preview the action, page regions, inset shots, and balloon positions.
In-panel balloon editing Move, resize, and edit speech balloons, adjust the tails, then recompose the comic.	Masked panel repair Mark a region and compare the repair candidate against the original before adopting it.
Production workbench Inspect panel generation records, batch jobs, parameter experiments, and effects layers.	Creative team and model settings Inspect agent roles and knowledge cards, and manage text, vision, and image connections.

Featured comic · We Need To Talk About Fifty-Five

A Chinese adaptation of the opening installment of qntm’s SCP Wiki serial. The finished chapter uses uneven panel sizes, inset close-ups, and dialogue placed inside the panels.

Read the full comic · Edit a balloon on artwork from this chapter · View full application screenshots

Paged layout preview

This preview uses existing chapter art with manually marked focal points and lettering exclusion areas. It demonstrates the compositor with the same artwork. End-to-end model generation was outside this check. Original chapter and credits

Multi-shot panel example · A hand detail and a brief memory sit inside separate main panels. The later location change gets its own panel. Read the adaptation · Validation record

Storyboard layout study · Original comic-page references and a same-art comparison of a uniform grid and a narrative layout. See the visual study

What makes this useful

• Six role-specific agents coordinate planning, storyboards, dialogue, artwork, review and layout.
• Edit the outline, panel descriptions and dialogue rather than restarting the whole chapter.
• Move/resize bubbles, edit their text, drag tail tips, and switch speech/thought/narration.
• Select a damaged region for repair, compare the result and explicitly adopt it. Compositing retains original source-panel pixels outside the mask.
• Plan full-width story beats and uneven detail/context rows. Recompose an existing chapter as pages or a vertical strip using the same artwork.
• Lock character, scene, and prop references for scene-level inheritance with per-panel overrides. Preview which later panels reuse a selected drawing before opting into a wider redraw.
• Keep character profiles and outfits, inspect per-panel request snapshots, and restore chapter versions.
• Configure text/vision and image providers independently. Use your own API services, the official Codex connector, or a combination.

These features support control and review; they do not guarantee visual consistency or make the project unique among comic tools.

Choose how models run

Open Creative Team & Settings (the application UI currently uses Chinese labels: 创作团队与设置).

Setup	Text & vision panel	Image panel	What you do
Official Codex connector	Codex	Codex	Sign in to ChatGPT, verify capabilities, and enable automatic calls
Automated providers	API	API	Supply separate base URLs, model names and your own keys
Mixed	Either	Either	Use an API for writing and ChatGPT for images, or the reverse

Text/vision uses OpenAI-compatible /chat/completions; visual review needs an image-capable model. Image generation uses /images/generations, reference-based generation uses multipart /images/edits. Masked editing also needs an editing-capable service. Native vendor-specific protocols need separate adapters.

Keys stay in the backend computer's .env and are not returned by the configuration endpoint. Settings writes are restricted to local access. You can keep an existing key by leaving its field empty, or explicitly clear it. Saving settings does not send a paid test or verify your provider's quota.

Manual generation handoffs are retired in v2.5.1. Existing tasks can migrate to an automatic provider; editorial approval is still available. This project does not extract session credentials or provide model access/credits. The GitHub Pages demo uses pre-generated artwork and makes no AI calls. The complete app runs locally.

Run locally

Requirements: Python 3.11+, Node.js 20+, Git. No GPU is required by the orchestration/editor itself; external or locally hosted models have their own requirements.

Windows one-click bootstrap: download the installer ZIP, extract it, then run Install-ComicCraft-Auto.cmd for the official connector setup, or Install-ComicCraft.cmd for the website only. It downloads tagged source and can install missing Python/Node through winget. Internet and any OS permission prompts still apply. It preserves existing local work. Details and limitations.

Or use Git on Windows:

git clone https://github.com/Seigi-91/comiccraft-ai.git
cd comiccraft-ai
powershell -ExecutionPolicy Bypass -File .\start.ps1

Open http://127.0.0.1:8000. The script creates a venv, copies .env.example if needed, installs dependencies and builds the frontend.

macOS / Linux:

git clone https://github.com/Seigi-91/comiccraft-ai.git
cd comiccraft-ai
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[dev]'
cp .env.example .env # only for a fresh installation; preserve an existing .env
npm --prefix frontend ci
npm --prefix frontend run build
python -m uvicorn app.main:app --host 127.0.0.1 --port 8000

Install a Chinese-capable font, such as Noto Sans CJK, and set COMIC_FONT_PATH if needed. A fresh install starts with an empty database. Click the home screen’s sample-import button to load an editable six-panel example without model calls. Re-import opens it without overwriting edits. Environment diagnostics are available there too.

Architecture

Layer	Implementation
Interface and API	React, TypeScript, and Vite power the workspace. FastAPI and Pydantic handle APIs and structured output validation
Orchestration and storage	LangGraph coordinates story, storyboard, dialogue, illustration, visual review, and lettering. SQLAlchemy and SQLite store task state, assets, versions, and generation records
Model connections	The local official Codex App Server calls cloud models using your own ChatGPT login. Compatible APIs can be configured for text and vision, image generation, and masked editing
Creative knowledge	Role skills and local lexical retrieval add relevant guidance to prompts. Retrieval ranks English terms and Chinese character bigrams within each role and records card IDs and hashes
Incremental revision	Each finished panel is checkpointed. Request fingerprints track prompts, reference content, and provider configuration. Review reuse depends on unchanged visual evidence
Image processing and lettering	Pillow fills panel frames with focal-aware crops when they fit the marked regions and saved balloons. It fits the full artwork when a crop would cut protected content and keeps balloon editing in source-image coordinates. The compositor also handles effects. Masked repair composites candidate pixels within the selected region and preserves original pixels outside it

Architecture · Official connector · Knowledge cards · Budgets and recovery

Limits and validation

• This is an early local-first app, not a public multi-user SaaS. There is no login, tenant isolation or per-user secret vault. Keep the full backend on loopback; set ENABLE_LOCAL_SETTINGS=false for administrator-managed remote installations and add appropriate security before exposing it.
• Character continuity uses profiles and reference images, not a newly trained consistency model. Face/hand/prop protection depends on visual-model judgment and may miss objects.
• ChatGPT style/size instructions are prompt constraints, not guaranteed inference seeds or sampling steps.
• Batch processing uses a local single-process scheduler. It is not a distributed production queue.
• Protocol tests use mock HTTP and deterministic handoff data. They do not prove image quality, real provider availability or successful paid generation.
• There is no full vector RAG; recent outlines and character records supply context.

python -m pip install -e '.[dev]'
python -m pytest -q
npm --prefix frontend ci
npm --prefix frontend run build

Learn more / contribute

Most detailed guides are currently in Chinese: model connections, setup, user guide, architecture, comparison, contributing.

Use issue templates for bugs, provider compatibility or feature proposals. Include redacted reproduction steps; never attach API keys or private projects. Small contributions to documentation, reproducible samples and compatibility reports are welcome. English UI localization and real provider verification are still open work.

If the workflow helps you, a voluntary GitHub Star makes it easier to find again. Successful installs and specific feedback are just as valuable.

License

Original code/docs: MIT. Third-party skills, fonts, music and components retain their own licenses; see THIRD_PARTY_NOTICES. See each demo’s attribution for source text and artwork details.

Creative skills and knowledge

Role skills define the handoff and checks for each stage. Local retrieval adds relevant cases and source-backed guidance. The workflow separates source facts from visual adaptation choices, preserves narrative beats, supports asymmetric panels and inset shots, and reserves in-panel space for separate speech balloons. Reference images guide identity and spatial structure while the current shot controls action and state. Review distinguishes continuity errors from editorial choices.

Role skills · Knowledge and sources · Workflow validation

Novel adaptation and storyboard review

1. Open 小说改编 · 导入原著, import TXT or Markdown, and select the excerpt, fidelity mode, panel count, and required facts. Import and chapter creation make no model calls.
2. Generate source-grounded character cards and review the outline. The workflow then prepares the storyboard and dialogue and pauses before image generation.
3. In 视觉参考与分镜草图, review the proposed page layout, scene state, and dialogue before drawing. Each panel can inherit up to three locked references by name, with manual overrides where needed. A page inset has its own artwork. An in-panel subshot is part of its parent drawing.
4. Recompose the pages or revise selected panels. Review the downstream references before choosing whether to redraw those panels too. Later chapters from the same imported novel can carry over character settings and the preceding completed chapter’s final shot.

This workflow is available on main. The linked v2.5.1 installer downloads the tagged release. In the live Wizard of Oz trial, only the two selected panels changed. A subsequent lettering save took about 1.9 seconds and added no model calls. See the validation record.

Original text and provenance stay separate from existing chapters. Bubble editing, masked repair, production tools and chapter rollback remain available. Each task retains its source/settings snapshot. Limits: 300,000 characters per import, 2MB files, 8,000 characters per excerpt and 200 excerpts. This is a per-excerpt workflow, not automatic whole-book adaptation or a guarantee of fidelity. Detailed guide.