refcontrol-flux-kontext-dataset

# Flux Kontext RefControl Dataset

This dataset was created for training **Flux Kontext RefControl LoRAs**.
It provides paired data of **control maps** (depth, pose, lineart, canny) and their **corresponding results** for reference-guided training.

---
## 📂 Dataset Structure

```text
dataset/
│
├── depth/
│ ├── control/ # depth maps
│ └── result/ # corresponding images
│
├── pose/
│ ├── control/ # pose skeletons / keypoints
│ └── result/ # corresponding images
│
├── lineart/
│ ├── control/ # lineart outlines
│ └── result/ # corresponding images
│
├── canny/
│ ├── control/ # canny edge maps
│ └── result/ # corresponding images
```

- Files in `control` and `result` share the same names.
Example: `depth/control/0001.png` ↔ `depth/result/0001.png`

---

## 🎯 Purpose

The dataset is designed to train LoRAs that:
- Preserve **identity** (faces, style, object details).
- Follow **control signals**: depth, pose, lineart, or canny edges.
- Enable consistent and controllable generation with **Flux Kontext** models.

---

## 📸 Data Source & Attribution

All images were sourced from **[Pexels](https://www.pexels.com/)** under the **CC0 license**.
For this dataset, we carefully selected **photo series** where:
- The **object or person remained the same**,
- But **pose, position, or composition changed** across the sequence.

This approach ensures strong consistency for reference-based training while enabling meaningful variation for control tasks.

🙏 **A huge thank you to the talented photographers on Pexels** for sharing their work openly and making this dataset possible.

---

## 🔗 Related LoRAs

- **[Depth Reference Fusion LoRA](https://huggingface.co/thedeoxen/FLUX.1-Kontext-dev-reference-depth-fusion-LORA)**
Trigger: `redepthkontext`
Preserves identity while following a depth map.

- **[Reference Pose LoRA](https://huggingface.co/thedeoxen/refcontrol-flux-kontext-reference-pose-lora)**
Trigger: `refcontrolpose`
Uses a pose skeleton to transfer style/identity to a new pose.

- **[Reference Lineart LoRA](https://huggingface.co/thedeoxen/refcontrol-flux-kontext-reference-lineart-lora)**
Trigger: `refcontrollineart`
Controls proportions with lineart while keeping style.

- **[Reference Canny LoRA](https://huggingface.co/thedeoxen/refcontrol-flux-kontext-reference-canny-lora)**
Trigger: `refcontrolcanny`
Uses canny edges for precise contour control.

---

# Flux Kontext RefControl 数据集
本数据集专为训练**Flux Kontext RefControl LoRA（低秩自适应）**而打造，提供了**控制映射（control maps）**（涵盖深度图、姿态骨骼、线稿轮廓、Canny边缘图）与其**对应生成结果**的配对数据，用于参考引导式训练。

---
## 📂 数据集结构
text
dataset/
│
├── depth/
│ ├── control/ # 深度图
│ └── result/ # 对应生成图像
│
├── pose/
│ ├── control/ # 姿态骨骼 / 关键点
│ └── result/ # 对应生成图像
│
├── lineart/
│ ├── control/ # 线稿轮廓
│ └── result/ # 对应生成图像
│
├── canny/
│ ├── control/ # Canny边缘图
│ └── result/ # 对应生成图像

`control` 与 `result` 文件夹内的文件命名完全一致。示例：`depth/control/0001.png` ↔ `depth/result/0001.png`

---
## 🎯 设计目标
本数据集旨在训练可实现以下功能的LoRA：
- 保留**身份一致性**（人脸、图像风格、物体细节）；
- 遵循**控制信号**：深度、姿态、线稿或Canny边缘；
- 支持使用**Flux Kontext**模型实现稳定可控的图像生成。

---
## 📸 数据来源与署名声明
所有图像均源自**[Pexels](https://www.pexels.com/)**平台，并遵循**CC0许可协议**。本数据集精心筛选了符合以下条件的**摄影序列**：
- 主体（物体或人物）保持一致；
- 但序列内的姿态、位置或构图存在变化。

该设计确保了参考式训练所需的强一致性，同时为控制任务提供了富有意义的变化空间。

🙏 衷心感谢Pexels平台上的优秀摄影师们无偿分享作品，为本数据集的构建提供了可能。

---
## 🔗 相关LoRA
- **[深度参考融合LoRA](https://huggingface.co/thedeoxen/FLUX.1-Kontext-dev-reference-depth-fusion-LORA)**
触发词：`redepthkontext`
功能：在遵循深度映射的同时保留身份一致性。
- **[参考姿态LoRA](https://huggingface.co/thedeoxen/refcontrol-flux-kontext-reference-pose-lora)**
触发词：`refcontrolpose`
功能：利用姿态骨骼将风格与身份迁移至新姿态。
- **[参考线稿LoRA](https://huggingface.co/thedeoxen/refcontrol-flux-kontext-reference-lineart-lora)**
触发词：`refcontrollineart`
功能：通过线稿控制图像比例，同时保留原有风格。
- **[参考Canny边缘LoRA](https://huggingface.co/thedeoxen/refcontrol-flux-kontext-reference-canny-lora)**
触发词：`refcontrolcanny`
功能：借助Canny边缘实现精准的轮廓控制。