gitmyhub

ComfyUi-Untwisting-RoPE

Python ★ 90 updated 3d ago

Training-free style transfer for DiT models.

A ComfyUI plugin that applies style transfer between images using the Untwisting RoPE technique, no extra training required. Works with DiT-based image generation models like Z-image and Anima.

PythonComfyUIsetup: moderatecomplexity 2/5

This repository is a plugin for ComfyUI, a node-based interface used for generating images with AI. The plugin adds a specific image generation technique called Untwisting RoPE, which allows style transfer between images without any additional training. Style transfer means taking the visual style of one image and applying it to the content of another.

The technique works with a category of AI image models called DiT models (Diffusion Transformer models). These are a newer class of image generation architectures. The plugin currently supports two specific models: Z-image (and its faster turbo variant) and Anima.

Installation is straightforward for anyone already using ComfyUI. You clone this repository into the custom nodes folder inside your ComfyUI installation, then restart the application. Once installed, the plugin shows up as new nodes you can connect in your workflow. The repository also points to a companion node that handles image scaling, which is required for the workflows to function correctly.

The README is brief and points to three external resources for getting started: example images showing what the technique can produce, pre-built workflow files you can load into ComfyUI, and a documentation file explaining what each parameter in the workflow controls. The README does not go into the underlying theory in depth; for that, it links to the original research paper on which this plugin is based.

This is a niche tool aimed at ComfyUI users who want to experiment with style transfer on DiT-based models without having to train or fine-tune anything.

Where it fits