The Modular Prompt Generator

The Modular Prompt Generator is a stand-alone application I created to generate structured AI prompts using a blend of original training descriptions and modular elements. I created this generator to improvise quickly while generating on the clock—rapidly experimenting with prompts until I strike gold, generating variations, and continuing the process. This was inspired by John Cage’s use of chance, and Brian Eno’s Oblique Strategies cards.

Quick Start Guide

1. Set the PRMPT CAT balance Knob

100% → Uses an Original Training Prompt* (unaltered).
0% → Builds a new prompt using Modular Categories:
- Subject (main focus)
- Details (descriptive elements)
- Environment (setting and atmosphere)
- Mood (emotional tone)
Anywhere in between → Blends structured descriptions with modular components.

The Modular Categories are directly extracted from the 37 original prompts. Instead of introducing AI-generated substitutions, this ensures that each generated prompt remains entirely within the wording of the LoRA training dataset—just rearranged into structured components. This approach keeps the LoRA fully engaged while preventing the global AI model from pulling from its own resources. This is my way of being an AI purist—ensuring that the generated prompts stay true to the LoRA’s trained vocabulary and visual intent.

Each time a prompt is generated using Modular Categories, one phrase from each category is selected, ensuring variety while keeping the language faithful to the original dataset.

2. Enable or Disable Randomization

Checked → Selects one of the 37 prompts in a random order each time.
Unchecked → Cycles through the 37 prompts in sequential order, moving to the next prompt with each generation.

This means that unchecked mode ensures you methodically work through the dataset one prompt at a time, while checked mode introduces more variation by selecting prompts randomly.

3. Generate a Prompt

Click "Generate Prompt" → Creates a structured AI prompt.
Click "Variations" → Generates alternate versions of the last prompt.

Ways to Experiment

Compare extremes → Generate a prompt at 100% and then at 0% to see the difference.
Blend structure & randomness → Use 50-75% for a mix of original and modular elements.
Lock vs. Randomize → Cycle through the 37 prompts in order or select them randomly by toggling the Randomize Original Prompt checkbox.
Refine with Variations → Adjust the emphasis of a prompt without changing its subject.

How It Works

1. The PRMPT CAT balance Knob

100% Uses the original description exactly as written, preserving its structure.75% Keeps most of the original prompt but swaps out a few descriptive elements.50% Blends original elements with modular substitutions from the same dataset.25% Mostly modular but retains some of the original structure.0% Fully replaces the original prompt with categorized phrases taken directly from the 37 training prompts.

Each prompt consists of multiple sentences. The slider controls how many sentences are replaced with modular elements.

At 100%, the generator selects a complete, unmodified prompt from the training dataset. As the slider decreases, portions of the original prompt are swapped with phrases directly taken from the dataset’s categorized components—but never with new AI-generated language. This ensures that the LoRA’s vocabulary remains intactwhile offering structured variations.

2. The Variations Button: Restructuring Emphasis

AI image generators prioritize the first sentence of a prompt, treating it as the most important information, while details at the end have less influence on the final output. Because of this, the order of information within a prompt significantly impacts the generated image—even when the content remains the same.

Clicking "Variations" doesn’t introduce new elements; instead, it rearranges the existing details, shifting emphasis within the scene. By restructuring the order of the subject, details, environment, and mood, the AI interprets the prompt differently, allowing for subtle changes in composition, lighting, or focal point—without altering the core description.

For example, take the original prompt:
"A weathered neon motel sign stands in front of a row of modest, yellow-painted rooms. The faded lettering and chipped paint hint at years of roadside wear. The setting sun casts long shadows, emphasizing the quiet dissolution of roadside commerce."

A variation might shift the details forward:
"Faded lettering and chipped paint hint at years of roadside wear. A weathered neon motel sign stands in front of a row of modest, yellow-painted rooms. The setting sun casts long shadows, emphasizing the quiet dissolution of roadside commerce."

Both describe the same scene, but the AI will now prioritize texture and wear instead of the motel sign itself.

By using Variations, you can explore how different visual aspects take precedence in the AI’s interpretation—without changing the fundamental elements of the scene.

3. What Are the 37 Original Training Prompts?

At 100% Influence, the generator selects one of 37 fully written descriptions from the original AI training dataset. Each of these prompts is:

A complete scene description from an 8x10 film photograph.
Structured with a clear subject, environment, and mood.
Designed to guide the AI toward a specific visual interpretation.

These prompts were written specifically to train a LoRA model based on these 37 images. The LoRA does not generate images itself—instead, it modifies how the main Flux model interprets prompts.

To ensure that the generated images remain within the LoRA’s intended style, this generator does not introduce outside AI-generated language. Instead, it selects only from pre-existing descriptions or modular elements directly taken from the dataset.

Additionally:

If randomization is enabled, a different prompt is selected randomly from the 37 available options.
If randomization is disabled, the generator moves through the 37 prompts in order, selecting the next one in sequence with each click.

4. How Were the Modular Categories Created?

Each Modular Category—Subject, Details, Environment, Mood—was extracted directly from the 37 training prompts by identifying recurring descriptive phrases. Instead of using full descriptions, the generator breaks them down into interchangeable elements, preserving their structure while allowing for new combinations.

Each time a modular prompt is generated, one element from each category is selected, ensuring that every output is a structured combination of pre-existing, LoRA-trained phrases.