6 Limitation and Conclusion

本文为论文《Wan 2.1：开放先进的视频生成基座模型》「论文原文」板块的对应章节，按 arXiv 官方全文逐字保留，层级与「中文全文翻译」板块一一对应。

Limitation. This study presents Wan, a foundational video generation model that has achieved notable advancements across multiple benchmarks. Specifically, significant improvements have been observed in motion amplitude (e.g., sports, dance) and instruction-following capabilities. Nevertheless, several limitations persist, which should be addressed to realize more potential of Wan. First, preserving fine-grained details in scenarios involving large motion remains a challenge for our method, as well as a broader issue within the field of video generation. Substantial efforts are required to improve the fidelity of videos with large-scale motion dynamics. Second, the computational cost associated with large-scale models remains prohibitive. Inference on a 14B model currently requires approximately 30 minutes on a single head-end GPU without additional optimization. To democratize video generation as a universally accessible AI tool, further research into efficiency and scalability is essential. Finally, there is still a lack of domain-specific expertise. As a foundational video model, we are committed to enhancing Wan’s general capabilities, yet performance in specific localized scenarios, such as education and medicine, may be insufficient. We aim to address this limitation by open-sourcing our latest models and fostering community-driven development across diverse specialized domains.

Conclusion. In this report, we publicly release Wan, our latest video model, which establishes a new benchmark for video generation. We provide a comprehensive overview of the architectural design of our Wan-VAE and DiT models, including detailed insights into their training approach, data curation processes, evaluation methods, and empirical results. Furthermore, we meticulously analyze our data preprocessing pipeline and the strategic adaptations implemented to optimize model training. This holistic approach is designed to drive advancements in the domain of video generation. We also extensively investigate downstream applications, such as image-to-video generation, video editing, and personalized video generation, to demonstrate the versatility and practical utility of Wan across diverse scenarios. In addition to open-sourcing the 14B model, we explore the feasibility of leveraging smaller-scale models for efficient video generation. Notably, our 1.3B model not only achieves competitive performance compared to larger counterparts but also enables seamless inference on consumer-grade GPUs, significantly enhancing accessibility and practicality for content creators. Looking ahead, we plan to focus on scaling both data and model architectures to tackle the most pressing challenges in video generation. Our ongoing efforts aim to provide the research community with more robust and versatile tools for video creation, fostering innovation and broader adoption in this rapidly evolving field.