---
license: cc-by-nc-sa-4.0
task_categories:
- image-to-image
language:
- en
- zh
- ja
tags:
- diffusion
- artifacts
size_categories:
- 10K<n<100K
dataset_info:
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': gt
          '1': output
          '2': raw
  splits:
  - name: train
    num_bytes: 23620669760
    num_examples: 60000
  - name: test
    num_bytes: 11118639211
    num_examples: 28896
  download_size: 34310079396
  dataset_size: 34739308971
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---

# Diffusion-Degradation Restoration Dataset (DDRD)
# 扩散退化恢复数据集

DDRD is a specialized collection designed to study and mitigate degradation phenomena occurring during diffusion-based image processing, specifically focusing on artifacts introduced during Color Enhancement or Color Transfer tasks.

DDRD 是一个专门用于研究和修复扩散模型处理过程中退化现象的数据集，聚焦于颜色增强或颜色迁移任务中产生的伪影。


## 1. Overview | 概览

The dataset simulates quality loss in high-speed diffusion inference by utilizing the Flux.1-kontext model, combined with block cache optimizations and distillation acceleration. It captures complex fidelity loss such as color drifts and structural inconsistencies.

该数据集通过 Flux.1-kontext 模型，结合 block cache 优化与蒸馏加速技术，模拟了高速扩散推理过程中的质量损耗。它捕捉了复杂的保真度损失，如颜色漂移和结构不一致性。

注：原图数据来源于 Pexels 的开源数据集，感谢它们对开源的贡献。

- [x] 开源 512 分辨率的数据集 | Release 512-resolution Datasets

- [x] 开源 1024 分辨率的数据集 | Release 1024-resolution Datasets

## 2. Dataset Structure | 数据结构

The dataset contains 29,636 triplets (Total: 88,908 images).
数据集包含 29,636 组三元组（共计 88,908 张图片）。

| Split | Triplets | Description |
| :--- | :--- | :--- |
| Train | 20,000 | Training set for restoration models |
| Test | 9,636 | Benchmark for evaluation |

### Data Fields
- raw: Simulated input for the diffusion process (Distorted). / 模拟扩散输入图（带退化）。
- gt: Pristine Ground Truth representing the ideal target. / 原始高质量真实目标图（无损）。
- output: Final image from the accelerated diffusion model, containing artifacts. / 经加速模型输出的、带有模拟伪影的图像。

```text
.
├── train/
│   ├── metadata.jsonl  # 20,000 triplets
│   ├── raw/            # Distorted inputs (low-step, block cache drift)
│   ├── gt/             # High-quality ground truth
│   └── output/         # Reference outputs from specific pipelines
└── test/
    ├── metadata.jsonl  # 9,636 triplets
    ...
```

## 3. Technical Implementation | 技术细节

DDRD models the fidelity loss from modern generative pipelines:
DDRD 模拟了现代生成管线中的保真度损失：

- **Model Used | 使用的模型**: Flux.1-kontext.
- **Acceleration Scheme | 加速方案**: Distillation methods are employed to simulate losses generated during low-step inference. / 采用蒸馏方法模拟低步数 (Low-step) 推理产生的损失。
- **Optimization Mechanism | 优化机制**: Integrated block cache logic to model color drifts and structural inconsistencies occurring during state-managed inference. / 引入 block cache 逻辑，模拟状态管理推理过程中的颜色偏移与结构不一致性。

## 4. Usage | 使用方法

```python
from datasets import load_dataset

# Load the dataset | 加载数据集
dataset = load_dataset("ModelMoe/DDRD20K")

# Access a sample | 访问样本
# Keys: 'raw', 'gt', 'output'
sample = dataset["train"][0]
sample["raw"].show()
```

## 5. License | 许可协议
This dataset is licensed under CC BY-NC-SA 4.0.

Commercial use is strictly prohibited. This includes:
Training commercial models.
Integration into paid services or products.

本数据集采用 CC BY-NC-SA 4.0 协议。严禁商业用途，包括但不限于：
训练商业化模型。
集成至付费服务或商业产品中。