DreamStyle: A Unified Framework for Video Stylization

TL;DR

DreamStyle is a versatile framework that unifies multiple video stylization methods, supporting text, style image, and first-frame guidance, and improves style consistency and video quality through a specialized training pipeline.

Contribution

It introduces a unified video stylization framework supporting multiple conditions and a data pipeline, outperforming existing methods in style consistency and quality.

Findings

Supports three stylization conditions: text, style image, first frame.

Uses LoRA with token-specific matrices for training.

Outperforms competitors in style consistency and video quality.

Abstract

Video stylization, an important downstream task of video generation models, has not yet been thoroughly explored. Its input style conditions typically include text, style image, and stylized first frame. Each condition has a characteristic advantage: text is more flexible, style image provides a more accurate visual anchor, and stylized first frame makes long-video stylization feasible. However, existing methods are largely confined to a single type of style condition, which limits their scope of application. Additionally, their lack of high-quality datasets leads to style inconsistency and temporal flicker. To address these limitations, we introduce DreamStyle, a unified framework for video stylization, supporting (1) text-guided, (2) style-image-guided, and (3) first-frame-guided video stylization, accompanied by a well-designed data curation pipeline to acquire high-quality paired video data. DreamStyle is built on a vanilla Image-to-Video (I2V) model and trained using a Low-Rank Adaptation (LoRA) with token-specific up matrices that reduces the confusion among different condition tokens. Both qualitative and quantitative evaluations demonstrate that DreamStyle is competent in all three video stylization tasks, and outperforms the competitors in style consistency and video quality.