Two stages, two specialists
Language generation and image generation are kept as separate stages — the story model decides what happens, the image model decides how it looks, and neither has to do the other's job.
The Panel Press · room 10 of 11
ComicCrafter AI
One little idea. A whole comic, panel by panel.
ComicCrafter AI turns a one-line idea into a comic, panel by panel: a locally hosted LLM writes the story, a fine-tuned image model paints the panels, and a Gradio interface ties the press together.
It was built collaboratively with Nishaant Singh — the repository lives under owner Saurabh Kumar's account.
the press queue — an idea slip and its empty panel rack
idea slip · “a robot learns to paint by watching the rain”
panel 1painted by the press
panel 2painted by the press
panel 3painted by the press
panel 4painted by the press
the rack is shown empty on purpose — real panels come from the local pipeline; the demo video is the proof
The actual problem
AI image generators produce single images on demand; a comic needs narrative structure — intro, storyline, climax, moral — before any panel can be drawn, and the panels have to belong to the same story.
The engineering split is deliberate: keep story generation and image generation as separate, swappable stages so each model does the job it is good at, and tie them together with an interface a person can actually steer.
The exploded machine
A one-line idea feeds a locally hosted gemma-2-2b-it model (served through LM Studio's OpenAI-compatible API), which writes a structured story with intro, storyline, climax and moral.
Idea slip — one line in
The operator types a one-line story idea; that slip is the whole input.Writer head — local gemma-2-2b-it via LM Studio
A locally hosted model (OpenAI-compatible API) writes a structured story: intro, storyline, climax, moral.Image press — fine-tuned Stable Diffusion / FLUX.1-schnell via ComfyUI
The structured story drives the image model, which paints the panels — story and image generation stay separate stages.Panel rack — the comic assembles
Panels land in the rack in story order, one per beat.Gradio console — the operator's desk
A Gradio interface ties the whole press together, running entirely locally.
collaborator plate — built with Nishaant Singh · repository owner Saurabh Kumar
watch the demo videoThe structured story then drives the image stage: a fine-tuned Stable Diffusion / FLUX.1-schnell model served through ComfyUI paints the panels.
A Gradio interface is the operator's console — one flow from idea slip to panel rack, running entirely locally.
Engineering decisions
Local-first generation
The story model runs locally through LM Studio's OpenAI-compatible API — the press works without any cloud key.
Structured stories, not freeform text
Requiring intro / storyline / climax / moral structure gives the image stage panel-worthy beats instead of prose to guess at.
Gradio as the press console
A Gradio interface keeps the whole pipeline operable in one place while the models behind it stay swappable.
Proof — verified results
- Story model
- gemma-2-2b-it via LM Studio (OpenAI-compatible API)repo
- Image model
- fine-tuned Stable Diffusion / FLUX.1-schnell via ComfyUIrepo
- Story structure
- intro · storyline · climax · moralrepo
- Interface
- Gradiorepo
- Proof of work
- demo video (Google Drive)verified live
- License
- MITrepo
every number here carries its source — my resume's figures and the synthetic-fixture results are marked for what they are