(stable version) (the model outputs text in English) (tested on local multimodal models from the Qwen family)
StoryChain Lab is an experimental environment for building narratives step by step using local multimodal models.
The program is not designed as a “book generator.”
Instead, it guides the model through a sequence of stages:
image analysis,
atmosphere and world interpretation,
character construction,
conflict creation,
plot outlining,
final story generation.
Each stage becomes context for the next one.
This allows the user to observe how the model develops a narrative layer by layer instead of generating everything in a single pass. It functions more like a narrative laboratory.
What CAN this tool do?
analyze images using multimodal models,
build narrative context progressively,
preserve continuity between consecutive stages,
generate worldbuilding, characters, and conflict drafts,
create experimental stories based on a single image,
allow the user to manually edit every stage,
run entirely locally through LM Studio.
The most important aspect of the project is the ability to observe the model’s “narrative reasoning” process across multiple steps.
In practice, this means the user can see:
how the model interprets an image,
how it develops motifs and themes,
how it builds relationships between elements of the world,
where it maintains coherence,
and where it begins to lose consistency.
What this tool is NOT
This is not:
a professional novel-writing engine,
a stable long-term memory system,
an AI that understands storytelling like a human,
a perfect story generator,
a system that guarantees logical consistency.
Local models still have limited contextual memory and limited narrative understanding.
They can produce highly interesting fragments, but they may also:
forget subtle details,
mix information between stages,
alter character personalities,
lose cause-and-effect relationships,
create inconsistencies in later parts of the story.
Most issues do not appear in the main plot structure, but rather in smaller narrative details.
Experimental Nature of the Project
StoryChain Lab is an experimental project.
The goal is not to replace the writer, but to demonstrate:
how a model can be guided through multiple stages,
how to build a narrative pipeline,
how contextual flow works between stages,
where models perform well,
and where they begin to lose coherence.
This tool is best treated as:
a narrative laboratory,
a sandbox for prompt engineering,
an experiment with local models,
a demonstration of modern LLM capabilities.
The most interesting part of this project is not that the model can write a perfect story.
What makes it fascinating is that, despite its many limitations, it can still preserve a surprisingly high degree of coherence across multiple stages of generation.
How to Use It
Download the .zip archive and extract it.
You must have:
LM Studio installed,
a local multimodal model,
and Python correctly installed.
Load the multimodal model inside LM Studio. Then start the local server in LM Studio:
click the [Developer] icon,
then click Start Server.
Next:
run install.bat to install the required dependencies,
then run start.bat.
A simple GUI will appear where you can load a test image. Then go through each analysis and generation stage from Stage 1 to Stage 6. Wait until the model fully completes each stage before moving to the next one. The [Save] button allows you to save the generated text.
VERY IMPORTANT: while generating text, do not click on previous stages. If you notice that the local model from the Qwen family has entered a loop (this most often happens in Stage 6), simply save the text and exit the environment. You may edit the text, but do so during the summary stages.