How Fable’s 2023 Showrunner Experiment Generated South Park-Style Scenes

In July 2023, Fable Studio demonstrated a system for generating animated scenes in a simulated South Park-like world. A user can provide a story direction, while software assembles characters, dialogue, visuals and voices. The demonstration is a research prototype, not an official South Park production or evidence that anyone could already make a polished season from one prompt.

The idea builds on an earlier Stanford and Google research project in which 25 generative agents inhabited a simulated town. Their architecture stored and retrieved memories and used reflections to influence behavior. I previously thought this could make game characters more responsive, with events in their simulated lives affecting later interactions.

Fable’s Showrunner project applies a related multi-agent idea to entertainment. The researchers describe simulated characters with backgrounds, memories, goals and relationships, and use a higher-level story prompt to guide events. These are computational character representations, not evidence that the agents are living people.

The clip shows a simulated classroom where a teacher and familiar cartoon characters discuss virtual worlds. The voices and animation evoke South Park, but the sample is an experimental imitation. Fable described its South Park work as research-only and said it would not release a tool for making public South Park episodes; the show’s creators did not authorize the demonstration.

Berman describes a pipeline using customized Stable Diffusion for imagery and language models and agent software for story generation. The research combines several components rather than relying on a single prompt alone. The creators supply a scenario and the resulting output still needs review; the demonstration does not establish a fully hands-off, broadcast-ready production process.

For the visual models, the paper describes training on approximately 1,200 character examples and 600 background examples associated with the show. Separate character and setting models allow the system to assemble scenes in its simulated world. These are the authors’ reported inputs for this prototype, not a claim that all visual rights were cleared.

For story and script, the researchers describe a system using language models, embeddings and character memories to maintain context. A user can specify a hero, setting, cast and broad story direction; the system then generates scenes and dialogue from those inputs and simulation data. The approach aims to turn character interactions into a coherent script.

The paper describes beginning with an idea, synopsis and major events across a simulated week, roughly three hours of simulation play time. Berman says the system selects material for an episode of about 22 minutes. That is the described research workflow and goal, not a guarantee that every run yields a complete, coherent episode of that length.

The workflow diagram separates the user’s direction from agent actions and simulation data. Those inputs feed language-model prompt chains and evaluation, followed by casting and scene construction. A staging system, camera and voice generation contribute to the final playback; multiple steps are involved beyond a short story prompt.

The researchers also propose plot patterns and “dramatic operators,” or DrOps, to help organize scenes and their connections. Devices such as foreshadowing or reversals can be incorporated into the prompt process. This is an attempt to address long-form structure, not proof that the system has solved continuity or dramatic writing.

The demonstration raises questions for writers, actors and illustrators. At the time of this video, Hollywood writers were on strike, with AI among the issues in labor negotiations. The project also raises questions about the use of existing characters, performances and training material; the video does not settle those rights questions.

I watched several generated examples and found them impressive, but I do not think they replicate the human storytelling and comic timing that make South Park work. Looking and sounding like a familiar show is different from consistently matching its writing. That is my judgment of the examples, not a controlled quality assessment.

I expect AI to disrupt entertainment and affect jobs, while also helping people work more efficiently and creating new kinds of work. Those are predictions, not outcomes established by this prototype. Fable said it was exploring other shows; at the time of the video, I could not try the system because its code had not been released. The broader ethical and creative questions remain open.