ai-2d-animation

Experiments in AI x 2D Animation

Creation that runs on its own.

After the tools, it became clear our real interest and talent was in making media ourselves, not making the tools for others. So we went deep on two things: data sources that could generate genuinely interesting content, and 2D animation pipelines. 2D animation is niche and specific, which is exactly why it was promising: by blending traditional animation technique with AI, we could make things purely generative video models can't, with real quality control. The method stayed consistent throughout: build discrete 2D states, layer and composite them, and let AI drive the switching between them.

Agentic Video Flow The foundation for the era. Give it a topic and it fetches the information in real time, summarizes it, writes a script, storyboards it, picks a voice, and generates backgrounds. Point it at any data source and it produces a finished short.

newsfishnow Our first channel built on the flow: a fish that reads the day's headlines, its mouth animated by the audio. The fish is a nod to SpongeBob, not the most original idea, but a good place to start. The account grew fast, about 7K followers in the first two weeks.

Madame Lenoir Next we raised the complexity of both the character and the scene. Madame Lenoir is a 2D character who speaks on demand, and you can change her clothes, her surroundings, and the music programmatically. Every layer of the image is generated with Nano Banana Pro, then masked and composited, and we generate her eye shapes for blinking and mouth shapes for talking, composited live to drive the animation.

Eternal Sun A meditation channel, and a chance to explore animating layers in 2D. Meditation is a daily practice, so a countdown timer on screen was meant to catch you mid-scroll and invite you to pause. The audience that found it was earnest and genuinely participatory in the comments, mostly 65 and older, which was a quiet, lovely surprise.

Character Creator (tech fishies) To make characters as rich as Madame Lenoir without the slow hand-work, I built an agentic system that generates and lets me quickly clean up every variation a character needs: eye shapes, mouth shapes, and more. With it I could spin up two characters fast and compose them into one scene, with a camera that cuts between each character and a wide shot of both. It is the same idea throughout this era: build the Lego blocks, then combine them.

Stylization & line-boil studies Generative video has a tell: it is too smooth and too perfect to feel alive. So we break it on purpose. We drop the frame rate, separate the outlines from the backgrounds, process each on its own, and recombine them with stylization and line boil so the image breathes like hand-drawn animation. The payoff was a scene I love: a cozy mouse reading a bedtime story to its pet bee.