Generative AI · Public case study

ComicForge Local

A safety-aware text-to-comic engine that converts an original story prompt into structured episodes, panel plans, optional generated artwork and rendered comic pages.

Backend
FastAPI
Local planning
Ollama
Rendering
HTML and PNG
Governance
Safety and IP guardrails

Why this system exists.

Text-to-image demos often stop at a single prompt and do not address structured storytelling, character planning, repeatable rendering, content controls or dataset governance. ComicForge treats comic generation as a complete system rather than a one-shot image request.

User prompt → safety and IP filter → template mode or local LLM → structured comic JSON → optional panel-image generation → HTML preview → PNG comic-page compositor

What the public implementation demonstrates.

  • Produces structured comic episodes with character, scene and panel planning.
  • Supports template-based deterministic tests and local LLM inference through Ollama.
  • Offers multiple image backends, including local ComfyUI and optional external generation services.
  • Renders both browser previews and final PNG comic pages with fallback artwork when necessary.
  • Includes model benchmarking and dataset-manifest validation for future governed fine-tuning.

What I designed and built.

  • End-to-end story-to-comic architecture.
  • FastAPI backend, Pydantic schemas and local Ollama integration.
  • Safety, originality and intellectual-property filtering rules.
  • Panel-image backend abstraction and Pillow-based page composition.
  • Benchmarking, dataset governance and implementation roadmap.

Reliability, safety and responsible use.

Engineering decisions

  • Template mode enables tests without requiring an LLM, image service or GPU workflow.
  • The renderer can create a complete page even when generated panel artwork is unavailable.
  • The repository is scoped to original, work-safe fictional storytelling and responsible dataset handling.

Limitations

  • Character consistency and reference-sheet workflows remain roadmap items.
  • Model weights and generated media are intentionally not stored in the public repository.
  • The current repository prioritises backend and generation architecture over a polished consumer frontend.

The implementation stack.

Public and inspectable. The repository contains the implementation, documentation and setup guidance available for this project.
FastAPIOllamaPydanticPillowComfyUILocal LLMsStructured JSON