World Simulator: Queer Erotica and the Absurdity of AI Video Models That Promise the World

Technical Summary: World Simulator: Queer Erotica and the Absurdity of AI Video Models That Promise the World

Problem Statement
The paper addresses the growing marketing narrative surrounding advanced generative video models, which are increasingly pitched as "world simulators" capable of modeling infinite realities and the physical world with total coherence. The authors argue that this claim of completeness is fundamentally flawed because these models systematically exclude significant aspects of embodied human experience, specifically sexuality. Commercial large video models (e.g., Sora, Veo, RunwayML) enforce strict safety filters that prohibit the generation of explicit content, while foundational models are trained on datasets that have been systematically "cleaned" of such material. Consequently, these systems are structurally blind to the body and sexual acts. The core problem is the hubris of labeling these systems as "world simulators" when they cannot simulate one of the most foundational biological experiences, creating a gap between the "all-seeing" marketing claims and the reality of their training data limitations.

Methodology
The project employs a "Critical Creative AI" methodology, utilizing adversarial translation to probe the latent space and representational limits of a state-of-the-art Large Video Model (LVM).

• Model Selection: The authors utilized the Wan 2.1 VACE (Video All-in-One Creation and Editing) pipeline. While not explicitly marketed as a world simulator, it shares the diffusion transformer (DiT) architecture of such systems. Crucially, as an open-weight tool, it allowed the authors to bypass the rigid API censorship of commercial platforms to process sensitive inputs.
• Input Data: The source material consists of clips from Boys in the Sand (1971), a classic queer underground film by Wakefield Poole. Clips were selected specifically for their intricate motion and focus on interlocking bodies, features anticipated to challenge the model's reconstruction capabilities.
• Adversarial Translation Process: The workflow, implemented in ComfyUI, involved passing the erotic source clips through the video-to-video pipeline. The authors manipulated the generation process by restricting the denoising step counts to create a spectrum from abstraction to representation:
  • Low Steps (1–2): The output remains a noisy, hazy blur where kinetic energy and structural composition are preserved, but semantic content is unresolved.
  • Higher Steps (3–5): As the step count increases, the model attempts to force ambiguous shapes into known concepts. This results in a "collapse" where fleshy, sexual forms are hallucinated into banal, mundane, or abstract representations, such as kitchen appliances, strange architectures, or domestic objects.
• Installation: The resulting sequences were edited into four continuous loops displayed on miniature screens embedded on ESP32 boards, emphasizing the system's endless failure to make sense of the erotic input.

Key Contributions

1. Visualizing Structural Blindness: The work provides a tangible, aesthetic visualization of the limits of synthetic knowledge. By demonstrating how the system transforms intimate acts into mundane objects, it exposes the "unknown" concepts within the model's latent space.
2. Critique of the "World Simulator" Label: The project challenges the inflated claims of AI marketing by demonstrating that a system trained on filtered internet-scale video data cannot claim to simulate the "world" while remaining structurally blind to sexuality.
3. Critical Creative AI Intervention: The paper situates itself within a lineage of works (e.g., The Zizi Project, Learning to See) that use machine learning not to demonstrate artistic capability, but to critically reflect on the ideological structures and social implications of the technology.
4. Reframing Sensuality in AI: The authors question the utility of "simulation" as a goal, suggesting that the machine's failure to parse the body offers a "poetic friction" that is more honest than a perfect, rigid rendering.

Results
The generated outputs successfully demonstrated the model's inability to reproduce the semantic content of queer erotica.

• Abstraction to Hallucination: The video sequences showed a clear progression from "fleshy ambiguity" at low step counts to distinct, recognizable domestic objects at higher step counts.
• The "Collapse": The system did not merely censor the content (blackout or refusal) but actively hallucinated surreal alternatives. Erotic choreography was transformed into scenes involving kitchen appliances and abstract flesh, highlighting the model's reliance on "safe" priors when faced with unlearned concepts.
• Installation Impact: The multi-channel display on miniature screens created a loop where the viewer dwells on the suggestive motion before the model "corrects" it into the banal, visually emphasizing the gap between the input reality and the system's interpretation.

Significance and Claims
The paper claims that the pursuit of a total "world simulator" is a "fool's errand" motivated by spectacle and profit, as all models inherently possess blind spots due to training data biases and architectural compression.

• Modest Claims on "Completeness": The authors explicitly state they are not arguing for the inclusion of pornography in AI models to achieve "completeness." They note that pornography itself is a "false simulation" of desire.
• Artistic Shift: The significance lies in shifting the role of the artist from helping machines simulate the world more accurately to using these tools to create "poetic, sensual expansions" of material experience.
• Epistemological Limit: The work asserts that the failure of the machine to parse the body reveals the inherent limits of machine representation, challenging the notion that current AI can offer a coherent, totalizing replica of the known world. The project serves as a critical intervention encouraging a skeptical orientation toward AI marketing promises and a deeper reflection on the limits of what can be simulated.