---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- image-generation
- stable-diffusion
- flux
- sdxl
- prompt-dataset
- lora-training
pretty_name: Prompt Generation Dataset (10K Narrative for Flux & SDXL)
size_categories:
- 10K<n<100K
---

# Prompt Generation Dataset (10K Narrative for Flux / SDXL)

This dataset (`prompt_gen_final_10k.jsonl` and `prompt_gen_final_10k.csv`) was used to train and fine-tune image-prompt models such as [KavinduHansaka/Llama-3.2-1B-ImageGen](https://huggingface.co/KavinduHansaka/Llama-3.2-1B-ImageGen).

It contains **10,000 curated narrative prompt samples** designed for **image generation models** like **Stable Diffusion XL** and **Flux**.  
Unlike raw tag-based datasets, the `target` field provides **natural paragraphs (≈80–100 words)** that describe cinematic scenes with attributes (lighting, textures, mood, etc.) woven into the prose.

---

## Dataset Format

The dataset includes the following fields:

- `input` – short seed description or raw caption  
- `style` – original style string  
- `style_normalized` – normalized style category (cinematic, realism, fantasy, anime, etc.)  
- `negative` – raw negative tags (original)  
- `negative_target` – cleaned negative prompt text (separate field for conditioning)  
- `attributes` – additional properties (lighting, mood, quality, etc.)  
- `aspect_ratio` – target aspect ratio (original metadata, not in `target`)  
- `width` – image width in pixels (original metadata)  
- `height` – image height in pixels (original metadata)  
- `target` – **final narrative prompt** used for training (~80–100 words, cinematic style)  
- `_aug` – marker for augmented samples (used to reach 10k)

---

## Example

| input | style | style_normalized | attributes | negative_target | target |
|-------|-------|------------------|------------|-----------------|--------|
| "misty pine forest at dawn" | fantasy | fantasy | cinematic_lighting, detailed, vibrant_colors | blurry, low quality | *A misty pine forest awakens at dawn, where towering trunks fade into rolling fog. Soft cinematic lighting breaks through the haze, illuminating fine details of dew-laden branches and mossy trunks. Vibrant colors enhance the contrast between soft golden light and cool shadows, with natural exposure balancing the scene. The composition layers foreground, midground, and background for depth, producing a painterly fantasy landscape that feels both grounded and otherworldly.* |

---

## Usage

You can load the dataset directly with 🤗 Datasets:

```python
from datasets import load_dataset

# Load JSONL
ds = load_dataset("KavinduHansaka/prompt-gen-10k-flux-sdxl", split="train")

print(ds[0])