基于扩散模型的动画短片智能生成系统研究OA

With the rapid advancement of artificial intelligence generated content(AIGC)technologies,video generation based on diffusion models has achieved significant breakthroughs.However,existing methods still face common chal-lenges when generating long-duration video sequences,particularly regarding unstable character appearances and a lack of character identity consistency,making it difficult to realize end-to-end automated creation from textual scripts to high-quality videos.This paper designs and implements a diffusion model-based system capable of generating minute-length animated short films.First,the system leverages a large language model to automatically generate screenplays and char-acter profiles.It further proposes an innovative strategy for generating storyboard reference images by integrating IP-Adapter with Stable Diffusion XL.To address character consistency issues in video generation,we introduce a dual-stage feature injection and dynamic constraint balancing mechanism.Specifically,the final frame of the initial generated video is used as the reference for character appearance features.These features are then injected frame-by-frame into the latent space,accompanied by a linearly decaying adapter strength coefficient that enables a smooth transition from strong to weak constraints.This effectively balances character consistency with natural motion dynamics.We have constructed a complete,human-in-the-loop animation short film generation pipeline.The system not only validates the effectiveness of our proposed algorithms in mitigating character identity inconsistency but also provides a practical technical framework and solution for automated and interactive video content creation.