Is anyone else interested in building/fine-tuning open video models specifically for high quality 2D animation?

A computer science student specializing in AI and deep learning has issued a call to action for the open-source AI community: collaborate to build and fine-tune video models specifically for high-quality 2D animation. The post argues that while current video models like Veo, Kling, Seedance, Wan, and LTX produce acceptable animation-like outputs, they fundamentally fail to replicate true 2D animation motion. Instead of frame-by-frame animation with proper timing, spacing, and smear frames, these models generate 'puppet distortion' or 'real-life motion wearing an anime skin,' inheriting the physics of live-action footage rather than the stylized motion grammar of cel animation.

The student emphasizes that most video models are trained on realistic, live-action data, leaving 2D animation as an underserved niche. While private tools may exist for studios, they are unlikely to be released publicly. The solution, they argue, is collective open-source effort: shared datasets, training recipes, and community-funded fine-tuning runs. This would enable specialized models competitive in this domain, helping indie creators and small teams produce animation more viably without replacing animators. The post has sparked interest in the AI and animation communities, highlighting a gap in current generative AI capabilities.