hkust-c4g/echostyle

EchoStyle is a research project that changes the artistic style of a video, for example turning a normal video clip into something that looks like it was drawn in Ghibli animation, ink wash painting, oil painting, or Disney style 3D. It works by fine tuning an existing large video generation model called Wan2.2, adding small trained add on modules known as LoRA adapters, using a technique the authors call reverse data synthesis. The paper describing this work was accepted at the ECCV 2026 conference. The project trains two separate LoRA adapters instead of one: one adapter handles the earlier, noisier steps of the video generation process, which shape the overall structure of the output, and the other handles the later steps, which fill in fine detail. Splitting the work this way is meant to preserve more detail in the final stylized video. The project also supports a faster, distilled version of the base model for quicker generation at a small cost to quality, and its training code is built to scale automatically across many machines and graphics cards without extra setup. Out of the box, EchoStyle supports fourteen art styles, including Ghibli, several types of anime, watercolor, ink wash, Disney 3D, Pixar style 3D, minimalism, ukiyo-e woodblock printing, oil painting, and the visual style associated with director Makoto Shinkai. The authors note that prompts written in Chinese give the best results, since the model was trained mainly on Chinese language style descriptions. Running this project is demanding. Inference needs eight Nvidia GPUs with large amounts of video memory, and training needs at least eight, with sixteen recommended for best results, plus specific versions of Python, PyTorch, and a library called Flash Attention. Users download the base Wan2.2 model, an optional faster checkpoint, and the trained EchoStyle adapters, then run provided scripts to either generate stylized videos or train new style adapters.