mikebrave/Image-Prompt-Library

---
license: mit
language: en
tags:
- image-generation
- prompts
- midjourney
- stable-diffusion
- wildcards
- prompt-engineering
- sqlite
size_categories:
- 100K<n<1M
datasets:
- mikebrave/Image-Prompt-Library
---

# Image Prompt Library

A SQLite-backed library of 61,235 curated image-generation prompts with style variants, wildcard vocabularies, and structured metadata.

## Overview

We gathered prompts from across the internet — Midjourney, Stable Diffusion, and many other model communities — and consolidated them into a single structured database. After analysis we concluded that most prompts can be classified into roughly **6 prompt styles** (syntax families):

| Style | Description |
|---|---|
| `comma-separated` | Classic token lists: `subject, lighting, mood` |
| `booru-tags` | Space-separated tag format from imageboards |
| `enhanced-prompt` | Weighted/sectioned prompts with emphasis markers |
| `everyday-speech` | Natural language descriptions |
| `lisp-like` | S-expression structured prompts |
| `structured-fields` | Field-labeled prompts (`Subject:`, `Lighting:`, etc.) |

We removed model-specific coding (aspect ratios, style codes, etc.), then identified repeated subjects and styles and turned them into **wildcards** with curated value pools. The wildcard placeholders were substituted into prompts, creating a flexible template system.

Originally there were over 700,000 prompts. We performed a quality check and narrowed it down to ~61,000 high-quality prompts, with a **1,000-prompt superset** (`publish-curated-styles-1000`) designed for broad-coverage testing and evaluation.

## Status

**Work in progress.** Not all styles are filled out for every prompt — the plan is to complete those over time. New prompts, prompt styles, and wildcards will be added as they're found. Contributions are welcome.

### Current Stats

- **61,235** curated prompts
- **6** syntax families (prompt styles)
- **82,369** filled style templates
- **19** wildcard definitions with **588** total values
- **7** style profiles populated with generated content (using Gemma 4 and Qwen 3)
- **1,000** evaluation superset for testing

## Database Schema

The SQLite database includes these tables:

- `prompts` — Core prompt records with identifiers, concepts, and metadata
- `prompt_style_profiles` — Style/syntax family definitions
- `prompt_templates` — Style-specific positive/negative templates with wildcard placeholders
- `wildcard_definitions` / `wildcard_values` — Wildcard vocabularies and their value pools
- `prompt_wildcard_bindings` — Links prompts to their wildcard dependencies
- `prompt_sets` / `prompt_set_members` — Named subsets (e.g., evaluation sets, curated pools)
- `prompt_versions` / `prompt_template_versions` — Change tracking
- `style_generation_jobs` — LLM generation job tracking

## File

- `prompts.db` — The complete SQLite database (trimmed from 700k+ to ~61k prompts)

## License

MIT