CHAOS Command-Driven Image Editing Workflow with Smart Recoloring, Outfit Changing, and Object Editing via ICEdit Prompts

使用说明：

1，上传图片

2，设置大模型（fill，icedit lora,clip l,t5xxl,ae）

3，输入指令（如：让她衣服变成黄色，裙子变成黑色,戴上眼镜），运行生图

4，在第一次生图还行的前提下，固定种子，运行图片放大。

此工作流的核心是ICEdit lora,它是一个iclora逻辑下的延申lora,最早的iclora的官方论文就提到了他们要做指令控图的事情，但是似乎他们现在正全力在视频vace项目上，这个方向的项目暂由另一个团队研究发布了。现阶段这个逻辑下工作流生成的首图分辨率不高，指令控制率仍需提高，但是用一句复杂指令来精准控制图片的雏形似乎已经有了。所以这个工作流会持续的改进和更新。

论文名称：In-Context Edit: Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer

论文地址：https://arxiv.org/abs/2504.20690

项目地址：https://github.com/River-Zhang/ICEdit

官方ICEdit lora 的下载地址：https://huggingface.co/sanaka87/ICEdit-MoE-LoRA/tree/main

使用过程遇到问题可留言或入群(QQ群：247228975)免费咨询。

Instructions for Use:

1. Upload an image

2. Configure the large model (fill, icedit lora, clip l, t5xxl, ae)

3. Enter editing instructions (e.g., "change her clothes to yellow, the skirt to black, and add glasses"), then generate the image

4. Once a satisfactory initial image is generated, fix the seed and proceed with image upscaling

The core of this workflow is the ICEdit lora, an extended lora based on the iclora framework. The original iclora paper mentioned their intention to develop instruction-driven image editing; however, it appears they are now focusing heavily on video-related VACE projects, leaving this direction to be further developed and released by another team. Currently, the resolution of the initial images generated through this workflow is not very high, and the instruction control accuracy still needs improvement. Nevertheless, the ability to precisely edit images using complex instructions is beginning to take shape. Therefore, continuous improvements and updates to this workflow are expected.

Paper Title: In-Context Edit: Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer

Paper Link: https://arxiv.org/abs/2504.20690

Project Repository: https://github.com/River-Zhang/ICEditThe official ICEdit lora download address: https://huggingface.co/sanaka87/ICEdit-MoE-LoRA/tree/main

If you encounter any issues during use, feel free to leave a message or join the QQ group (247228975) for free consultation.