Make Pixels Dance: High-Dynamic Video Generation

Yan Zeng,Guoqiang Wei,Jiani Zheng,Jiaxin Zou,Yang Wei,Yuchen Zhang,Hang Li

2023-11-18

Abstract:Creating high-dynamic videos such as motion-rich actions and sophisticated visual effects poses a significant challenge in the field of artificial intelligence. Unfortunately, current state-of-the-art video generation methods, primarily focusing on text-to-video generation, tend to produce video clips with minimal motions despite maintaining high fidelity. We argue that relying solely on text instructions is insufficient and suboptimal for video generation. In this paper, we introduce PixelDance, a novel approach based on diffusion models that incorporates image instructions for both the first and last frames in conjunction with text instructions for video generation. Comprehensive experimental results demonstrate that PixelDance trained with public data exhibits significantly better proficiency in synthesizing videos with complex scenes and intricate motions, setting a new standard for video generation.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper attempts to address the challenges faced in generating high-dynamic videos (such as action-rich scenes and complex visual effects) in the field of artificial intelligence. Current state-of-the-art video generation methods mainly focus on text-to-video generation, and although they can maintain high fidelity, the generated video clips usually have limited action. Therefore, the paper proposes a new method—PixelDance, which is based on a diffusion model and combines image instructions of the first and last frames with text instructions to generate videos.

Specifically, the paper addresses the following aspects:
1. **Limitations of existing technologies**: Existing video generation technologies perform poorly in generating complex scenes and actions, especially when dealing with content outside the training data.
2. **Improving video quality**: By introducing image instructions of the first and last frames, PixelDance can generate higher quality and action-rich videos.
3. **User control**: Users can provide rough sketches as instructions using simple image editing tools, making the generated videos more in line with user expectations.
4. **Long-duration video generation**: PixelDance can generate continuous video clips and exhibits better temporal and spatial consistency in long video generation.
5. **Cross-domain generation capability**: Even if the training data lacks non-realistic style content (such as sci-fi, comics, and cartoons), PixelDance can still generate high-quality videos.

Overall, the paper aims to significantly enhance the quality and diversity of video generation by combining image and text instructions.