Summary: Feasibility Study on AI-Assisted Virtual Character Rigging

This report evaluates the practical efficacy of using AI agents—specifically Claude (utilizing Opus 3.5) and the Codex CLI—to automate the creation and rigging of Live2D-style virtual characters. The author concludes that currently, relying on AI for this process is significantly less efficient than traditional manual workflows, highlighting a "productivity paradox" where the labor cost of managing AI interventions far exceeds the cost of professional manual production.

#### Core Methodology and Technical Workflow
The author attempted to bridge image generation and rigging automation through a multi-modal AI stack:

1. Environment Setup: Integrated Node.js and npm to interface with the Codex CLI, allowing the AI to generate and retrieve character components via terminal-based prompts.
2. Structural Decomposition: Initial attempts to have the AI segment full-body illustrations into functional rigging parts failed, resulting in "chimera-like" anatomical distortions. This necessitated a shift to a manual, iterative structure:
   • Part Segmentation: The character was manually broken down into discrete layers (eyes, hair, facial features).
   • Mesh & Deformer Mapping: For each segment, the author utilized the AI to guide the creation of meshes and deformers. Parameters were manually defined for head rotation, eye blinking, and facial expressions.
3. Iterative Correction Loop: The core "automated" process relied on a feedback loop:
   • The current state of the character was periodically captured.
   • Codex was prompted to generate reference images based on correct anatomical proportions.
   • Claude acted as the intermediary to compare the deformed mesh against the reference, requiring the author to manually tweak coordinate placements to maintain consistency across rotations ($X, Y, Z$ axes).
4. Physics and State Logic: The author implemented "spring-like" physics for secondary motion (hair, ribbons, earrings). State transition rules were defined for facial expressions (e.g., smile vs. open mouth) to ensure non-overlapping parameters, preventing the model from conflicting expressions.

#### Challenges and Technical Bottlenecks

• Anatomical Inconsistency: AI-generated references lacked temporal/spatial consistency. Every time a new reference was generated for a different angle (e.g., $90^\circ$ rotation), the model struggled to maintain geometric continuity with previous frames.
• High Latency in Error Correction: Because the AI cannot directly manipulate the rigging software’s vertices, the user acted as an "operator" translating natural language instructions into precise geometric adjustments. This resulted in a total production time of over 40 hours, compared to an estimated 8 hours for an experienced professional.
• Integration Overhead: The "hidden cost" of utilizing AI (subscription fees plus the 40+ hours of manual labor) was estimated at 400,000 KRW in labor equivalent. The project demonstrated that for simple character rigging, AI tools currently create more "maintenance debt" than they resolve.

#### Conclusion
The experiment highlights that while AI can assist in the conceptualization of assets, the lack of a deep, native integration between LLMs and rigging software makes the current pipeline inefficient. The complexity of managing $n$ number of parameters ($p_1, p_2, \dots, p_n$) across varying angles while ensuring the integrity of the mesh surface requires human intervention that currently negates any speed benefits promised by automation.