What is hy-wu?

HY-WU is a research framework for AI powered image editing using text instructions. The core capability it demonstrates is transferring visual attributes between images, for example taking an outfit from one photo and placing it onto a person in a different photo, or transferring a face's identity while keeping the background and pose unchanged. The technical problem it solves is a longstanding tradeoff in AI image generation: customizing a model to handle a specific input, like a particular person's outfit, normally requires fine tuning the model on that input, which is slow and expensive. HY-WU instead generates lightweight adapter weights on the fly for each individual request, based on the input images and a text instruction. These adapters modify the image generation process without permanently altering the underlying model. This means each image editing request can be processed independently without any training or optimization step at the time of use. The README shows several example use cases, including combining clothing across different images, cosplay outfit transfer, transferring a face's identity onto another image, and generating detailed textures. The project reports human evaluation results claiming it performs close to some of the strongest closed source commercial image models available, while remaining open source. The framework is built to scale up to very large base models, and the main checkpoint requires multiple high end GPUs to run due to the model's size, specifically at least 8 GPUs with 40GB of memory each, or 4 GPUs with 80GB each. The project includes example code and an optional web based demo interface. You would use this if you are an AI researcher studying image editing, personalization, or parameter generation, or if you want to run text guided image editing that involves transferring visual elements between images. It is written in Python and was published by Tencent's Hunyuan research team.