improved-flux-prompts

## Experimental FLUX Prompt Dataset

## Overview
This dataset features a curated selection of prompts designed specifically for FLUX.1, an innovative text-to-image synthesis model. The prompts are crafted to produce high-quality, imaginative images by utilizing advanced prompting techniques and best practices.

## Dataset Improvements 🚀

🎉 We've completed an additional cleaning and curation process for this dataset. Here's what's new:

- ✨ Removed excessive term repetition
- 🔍 Enhanced prompt quality and diversity
- 📊 Improved overall dataset consistency

This refined version offers an even more valuable resource for FLUX.1 users, providing a versatile collection of prompts for your creative projects.

🔄 Ongoing Curation
While we've made significant improvements, our commitment to excellence continues. This dataset remains under curation, and we're always looking for ways to enhance it further. If we decide to implement any additional improvements, we'll upload a new version promptly.

For now, that's it! 🎨 Enjoy exploring this enhanced dataset and unleashing your creativity with FLUX.1!
## Example Result
We've put together a sample output grid to give you a visual representation of what the dataset can do. This grid features 9 images created from prompts in this dataset.




Every image in the grid features a watermark that displays its corresponding prompt ID, making it easy to reference.

## Creating a Dataset
The dataset was generated using a multi-step approach:

1. **Base Examples**: We began by choosing the top-performing prompts from CivitAI to serve as our foundation.

2. **LLM Enhancement**: The prompts underwent additional refinement and expansion with the help of the Hermes3 Large Language Model (LLM).

3. **FLUX Optimization**: The prompts were subsequently refined for FLUX.1, utilizing best practices and techniques recognized for delivering outstanding results with this model.

## Features of the Prompt
Every prompt in this dataset is crafted to encompass:

- Bright and intricate portrayals
- Distinctive artistic styles
- Guidelines for composition
- Details on lighting and color schemes - Indicators for mood and atmosphere
- Relevant technical details

## Data Quality Note
Although there have been attempts to refine and enhance the prompts by steering clear of Stable Diffusion-style prompting and eliminating references to LoRA names, it's possible that some noise might still linger in the dataset. Even so, the initial results have been quite promising, leading to the choice to share this dataset with the community. It's a good idea for users to stay mindful of possible inconsistencies and to fine-tune their prompts as necessary for their particular needs.

## System Message for LLM
The Hermes3 LLM was directed by a thoughtfully designed system message to guarantee top-notch, FLUX-optimized prompts. The main points of this system message are:

- Guidelines for crafting clear and thorough descriptions
- Emphasize the details of the image instead of the steps taken to create it.
- Focus on creating distinctive prompts that draw inspiration from examples without replicating them - Recommendations for including key elements such as subject, style, composition, lighting, color palette, mood, and technical specifications - Strategies for developing impactful prompts, including the use of artistic references, technical jargon, concept blending, and style fusion

You can find the complete system message in the `flux_system_message.txt` file located in this repository.

## Dataset Structure
The dataset is organized in a JSONL (JSON Lines) format, where each line corresponds to an individual prompt. Every entry includes:

- `id`: A distinct identifier for the prompt
- `prompt`: The specific text of the prompt

## How to Use
This dataset is designed for experimental purposes with FLUX.1 and comparable text-to-image models. These prompts can be utilized by researchers and developers to:

1. Create top-notch images
2. Explore effective prompting techniques
3. Evaluate the performance of the model

## Acknowledgements
First and foremost, we want to express our gratitude to Black Forest Labs for creating the innovative FLUX.1 models that made this project possible. Their groundbreaking work in text-to-image synthesis has opened up new frontiers in AI-generated art.

We also extend our thanks to the Flux Reddit community, which has served as an invaluable incubator for ideas and techniques. The discussions, experiments, and shared knowledge within this community have greatly influenced and improved our approach to prompt engineering.

Additionally, we want to give a shoutout to the CivitAI community for sharing the initial set of high-performing prompts that inspired this dataset. Their contributions have been instrumental in establishing a strong foundation for our work.

## License
MIT

### 实验性FLUX提示词数据集

## 概览
本数据集专为文本到图像合成模型FLUX.1打造，收录了经过精心筛选的提示词。这些提示词借助先进的提示工程技巧与行业最佳实践创作而成，旨在生成高质量、富有想象力的图像。

## 数据集优化 🚀
🎉 本数据集已完成额外的清洗与精选流程，新增优化内容如下：
- ✨ 移除了冗余的术语重复内容
- 🔍 提升了提示词的质量与多样性
- 📊 优化了数据集整体的一致性
本次优化后的版本可为FLUX.1用户提供更具价值的资源，包含多样化的提示词集合，助力您的创意创作。
🔄 持续精选工作
尽管已完成多项重要优化，我们对卓越品质的追求仍未停歇。本数据集仍处于持续精选迭代中，我们将持续探索进一步优化的路径。若后续推出额外改进版本，我们将及时上传更新。
目前本数据集已准备就绪 🎨 欢迎探索这款优化后的数据集，尽情释放您使用FLUX.1的创作灵感！

## 示例效果
我们制作了示例输出网格图，直观展示本数据集所能实现的效果。该网格包含9张由本数据集内提示词生成的图像。
网格内每张图像均带有水印，显示其对应的提示词ID，便于快速查阅引用。

## 数据集构建流程
本数据集通过多阶段流程生成：
1. **基础示例集**：我们首先选取CivitAI平台上表现优异的提示词作为数据集基础。
2. **大语言模型优化**：借助Hermes3大语言模型（Large Language Model, LLM）对提示词进行进一步的细化与拓展。
3. **FLUX专属优化**：随后结合FLUX.1的行业最佳实践与专属技巧，对提示词进行针对性优化，确保其在该模型上可输出卓越效果。

## 提示词特性
本数据集内的所有提示词均覆盖以下要素：
- 明亮且细节饱满的画面呈现
- 独具一格的艺术风格
- 构图规范指引
- 光照与色彩方案细节，以及情绪氛围设定指标
- 相关技术参数说明

## 数据质量说明
尽管我们已尝试规避Stable Diffusion风格的提示词写法、移除LoRA名称引用，对提示词进行了优化完善，但数据集内仍可能存在少量冗余噪声。尽管如此，首批测试结果已相当可观，因此我们决定将本数据集分享至社区。建议用户留意潜在的不一致问题，并根据自身需求对提示词进行必要的微调。

## 大语言模型系统提示词
本数据集使用Hermes3大语言模型时，依托精心设计的系统提示词来保障生成符合FLUX优化标准的高质量提示词。该系统提示词核心要点包括：
- 清晰且详尽的描述创作规范
- 强调图像细节而非图像生成流程
- 聚焦创作独特的提示词，可借鉴示例灵感但避免直接复刻
- 建议纳入主体、风格、构图、光照、色彩搭配、情绪氛围与技术参数等核心要素
- 打造高影响力提示词的策略，包括艺术参考引用、专业术语运用、概念融合与风格混搭
完整的系统提示词可在本仓库的`flux_system_message.txt`文件中查看。

## 数据集结构
本数据集采用JSON Lines（JSONL）格式组织，每一行对应一条独立的提示词。每条数据包含以下字段：
- `id`：提示词的唯一标识符
- `prompt`：提示词的具体文本内容

## 使用说明
本数据集专为FLUX.1及同类文本到图像模型的实验性应用设计。研究人员与开发者可利用这些提示词完成以下工作：
1. 生成高质量图像
2. 探索高效的提示工程技巧
3. 评估模型的性能表现

## 致谢
首先，我们谨向Black Forest Labs致以诚挚谢意，感谢其推出创新的FLUX.1系列模型，为本项目奠定了核心基础。其在文本到图像合成领域的突破性工作，为AI生成艺术开辟了全新的发展前沿。
同时，我们也要感谢Flux Reddit社区，这里是极具价值的创意与技巧孵化阵地。社区内的讨论、实验与共享知识，极大地影响并优化了我们的提示工程思路。
此外，我们还要感谢CivitAI社区分享的首批优质热门提示词，这些内容为本数据集提供了宝贵的创作灵感与坚实基础。

## 许可证
MIT