---
title: ImagePulse-StyleTransfer
canonical_url: "https://www.modelscope.cn/datasets/DiffSynth-Studio/ImagePulse-StyleTransfer"
md_url: "https://www.modelscope.cn/datasets/DiffSynth-Studio/ImagePulse-StyleTransfer.md"
repository: DiffSynth-Studio/ImagePulse-StyleTransfer
chinese_name: "图律脉动数据集-风格迁移"
last_updated: 2025-04-20
license: "Apache License 2.0"
storage_size: "703 GB"
downloads: 80678
stars: 26
---

# ImagePulse-StyleTransfer

> ImagePulse-StyleTransfer - DiffSynth-Studio 在 ModelScope 开源的数据集。本数据集由开源项目 ImagePulse 构建，可用于训练图像风格迁移能力。

DiffSynth-Studio/ImagePulse-StyleTransfer 是 ModelScope 魔搭社区上的数据集，存储大小 703 GB，采用 Apache License 2.0 许可。

- **Repository**: DiffSynth-Studio/ImagePulse-StyleTransfer
- **License**: Apache License 2.0
- **Storage size**: 703 GB
- **Downloads**: 80678
- **Stars**: 26
- **Last updated**: 2025-04-20

Source: https://www.modelscope.cn/datasets/DiffSynth-Studio/ImagePulse-StyleTransfer

---

# 图律脉动数据集-风格迁移

本数据集由开源项目 [ImagePulse](https://github.com/modelscope/ImagePulse) 构建，可用于训练图像风格迁移能力。

主要字段说明：

* `image_1`: 图像1，内容图片
* `image_2`: 图像2，风格图片
* `image_3`: 图像3，风格迁移图片（低质量中间结果）
* `image_4`: 图像4，风格迁移图片
* `editing_instruction`: 将图像1编辑成图像4所需的文本指令
* `reverse_editing_instruction`: 将图像4编辑成图像1所需的文本指令

数据集构建流程：

* 在数据集 [DiffusionDB](https://modelscope.cn/datasets/AI-ModelScope/diffusiondb) 中随机抽取提示词
* 用 [FLUX](https://modelscope.cn/models/black-forest-labs/FLUX.1-dev) 模型根据提示词生成图像1
* 用 [Qwen-VL](https://modelscope.cn/models/Qwen/Qwen2.5-VL-72B-Instruct) 模型删除提示词中与风格相关的部分，仅保留内容提示词
* 用[风格提示词模板](https://github.com/modelscope/ImagePulse/blob/main/pulse/processor/style.py)和内容提示词构造随机的风格迁移提示词
* 用 [FLUX](https://modelscope.cn/models/black-forest-labs/FLUX.1-dev) 模型根据风格迁移提示词生成图像2
* 用 [SDXL](https://modelscope.cn/models/AI-ModelScope/stable-diffusion-xl-base-1.0)、[IP-Adapter](https://modelscope.cn/models/AI-ModelScope/IP-Adapter)、[Union-ControlNet](AI-ModelScope/controlnet-union-sdxl-1.0-promax) 组装的模型进行风格迁移，启用 InstantStyle 得到图像3
* 用 [FLUX](https://modelscope.cn/models/black-forest-labs/FLUX.1-dev) 和 [IP-Adapter-FLUX](https://modelscope.cn/models/InstantX/FLUX.1-dev-IP-Adapter) 模型润色图像3，得到图像4
* 用 [Qwen-VL](https://modelscope.cn/models/Qwen/Qwen2.5-VL-72B-Instruct) 模型重新生成对偶编辑指令

数据样例：

|image_1|image_2|image_3|image_4|editing_instruction|reverse_editing_instruction|
|-|-|-|-|-|-|
|![](examples/1744027631796853160.png)|![](examples/1744027632264796657.png)|![](examples/1744027632584051560.png)|![](examples/1744027633022068239.png)|transform the image into a cartoon style with vibrant colors and a confident expression.|transform the image into a realistic portrait with a serious expression and subtle lighting.|
|![](examples/1744031879395794851.png)|![](examples/1744031879898078896.png)|![](examples/1744031880408740400.png)|![](examples/1744031880783716279.png)|transform the image to have a brighter, more colorful palette and a clear blue sky.|transform the image to have a more muted color palette and an overcast sky.|
|![](examples/1744050264122532325.png)|![](examples/1744050264640221066.png)|![](examples/1744050265127669284.png)|![](examples/1744050265613174226.png)|transform the style of the image to an anime illustration, change the jacket to red, and add a cityscape background.|transform the style of the image to a digital painting, change the jacket to black, and remove the cityscape background.|