Abstract
Interactive Character Editing (ICE) is a novel framework designed to streamline the creation and refinement of 3D game characters through natural language dialogue. By integrating Large Language Models (LLMs) with a semantic-guided parameter solver, ICE translates high-level user requests into precise adjustments of hundreds of in-game technical parameters. This approach enables an iterative, collaborative design process that significantly reduces the manual labor traditionally required for character customization.
Introduction
Modern Role-Playing Games (RPGs) offer extensive character customization, often involving hundreds of parameters ranging from facial bone structures to makeup details. However, manually adjusting these parameters is a time-consuming and labor-intensive task. While existing AI-driven "auto-creation" tools attempt to automate this process, they typically follow a single-round approach that lacks a refinement phase. If a generated character requires specific minor adjustments, these systems often necessitate a full regeneration, losing the progress made. Furthermore, direct parameter optimization frequently leads to anatomically impossible or "uncanny valley" results. ICE addresses these challenges by providing a multi-round dialogue interface for fine-grained, iterative character editing.
Methodology
The ICE framework bridges the gap between natural language and game engine parameters through two core modules:
Instruction Parsing Module (IPM)
The IPM utilizes an LLM (e.g., GPT-4) to manage the conversational flow. To maintain consistency across multiple rounds of editing, the module incorporates a Character Attribute Memory Bank that tracks the state of previously modified attributes. The IPM processes user input to generate a clean text prompt, determine editing strength, and provide a natural language response to the user.
Semantic-guided Low-dimension Parameter Solver (SLPS)
To translate text prompts into character adjustments, the SLPS employs three key components:
- Character Control Parameter Localizer: A RoBERTa-based model identifies the specific parameters relevant to a user's request, ensuring that localized edits (e.g., "pointy nose") do not affect unrelated facial features.
- Low-dimension Space (PCA): To maintain anatomical plausibility and improve optimization speed, bone parameters are projected into a 60-dimensional latent space using Principal Component Analysis (PCA).
- Neural Imitator: Since game engines are non-differentiable, a neural network is trained to mimic the game's renderer. This allows the system to use gradient descent to align character parameters with the CLIP embedding of the user's prompt.
Experiments and Results
The framework was evaluated using models from Justice Online Mobile and Naraka: Bladepoint, utilizing a 450-dimensional parameter space. A PCA-based prior was trained on 10,000 character sets derived from the CelebA dataset, while the Parameter Localizer was trained on a custom dataset of 10,000 instruction strings.
Performance Metrics
- Efficiency: ICE achieves a response time of approximately 9 seconds per round, significantly faster than existing methods like TADA, which can take hours.
- Alignment: ICE outperformed competing methods (DreamFace, TADA, T2P) with a higher CLIP score of 0.2699, indicating superior alignment between user descriptions and the resulting 3D models.
- User Satisfaction: In a study involving 100 participants, ICE was preferred 77.1% of the time, largely due to its ability to maintain semantic consistency during fine-grained edits.
Discussion and Future Work
A key innovation of ICE is the use of semantic masking, which prevents unintended changes to unrelated facial features. Additionally, by constraining optimization to a low-dimensional space derived from real human faces, the system ensures realistic results even with extreme prompts.
Despite its advantages, ICE faces limitations in optimization latency for real-time applications and a potential lack of diversity for highly stylized or non-human characters due to its reliance on human-face priors. Future research will focus on further reducing latency and expanding the system's generative capabilities to include a wider variety of artistic styles.
Summary updated