EchoStyle

EchoStyle: Unlocking High-Fidelity Video Stylization with Reverse Data Synthesis

ECCV 2026 · Published paper · arXiv:2606.25465 · Project page · Source code

📖 Introduction

EchoStyle is a high-fidelity video stylization framework built on the Wan2.2 video generation model. It fine-tunes the Wan2.2 14B DiT model using LoRA adapters trained with a novel reverse data synthesis pipeline, enabling faithful style transfer across diverse artistic styles — Ghibli, anime, ink wash, oil painting, ukiyo-e, Disney 3D, and more.

TL;DR: EchoStyle is a text-driven video-to-video stylization method for short and arbitrary-length videos. It introduces reverse data synthesis for building the V-Style20k paired dataset and combines init-follow-mode training with sliding-window inference to improve temporal consistency.

Key Features

  • 🎨 Multi-style video transfer — supports 14 artistic styles with a single framework
  • 🔄 Dual-LoRA architecture — separate LoRA models for high-noise (t=900-1000) and low-noise (t=0-900) timestep ranges, yielding better detail preservation
  • ⚡ Lightning inference — compatible with LightX2V distilled checkpoints for 10-step generation
  • 🖥️ Scalable training — auto-adapts to multi-node multi-GPU clusters via FSDP + Ulysses sequence parallelism

📦 Checkpoints

This repository contains the two LoRA checkpoints used by EchoStyle:

  • echostyle_high.pth: high-noise LoRA for timesteps t=900–1000
  • echostyle_low.pth: low-noise LoRA for timesteps t=0–900

For installation and inference instructions, see the EchoStyle GitHub repository.

📝 Citation

@inproceedings{li2026echostyle,
  title     = {EchoStyle: Unlocking High-Fidelity Video Stylization with Reverse Data Synthesis},
  author    = {Li, Huaqiu and Wang, Jiahao and Cai, Sijia and Sheng, Hualian and Deng, Bing and Ye, Jieping and Luo, Wenhan},
  booktitle = {Computer Vision -- ECCV 2026},
  series    = {Lecture Notes in Computer Science},
  volume    = {17064},
  pages     = {270--286},
  publisher = {Springer, Cham},
  year      = {2026},
  doi       = {10.1007/978-3-032-37196-6_15},
  url       = {https://doi.org/10.1007/978-3-032-37196-6_15}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for youchun/echostyle