LLM-Driven 3D Interactive Development in Practice: Recreating a Digital Statue from Concept to Delivery

With the deepening integration of multimodal large language models and code generation capabilities, the role of AI in software development is shifting from a code completion tool toward a more autonomous, end-to-end development agent. Recently, an intelligent programming tool released a new-generation model version that reportedly demonstrates stronger capabilities in 3D scene understanding and complex front-end engineering delivery. To evaluate this trend, we designed a challenging end-to-end test: have the large language model independently build a 3D interactive website centered on recreating the birth of the classic sculpture David.

1. Task Definition and High-Level Prompt Engineering

To push the model beyond merely writing a block of code, we deliberately avoided granular instructions and instead adopted a goal-oriented, high-level prompting strategy.

We provided the model with the following core requirements:

1. Primary objective: Build a Chinese-language interactive website for a general audience that explains how a sculpture emerges from raw stone.
2. Emotional goal: After browsing, visitors should not only gain historical knowledge, but also develop a new understanding of carving itself and feel compelled to stop and observe.
3. Creative latitude: The narrative angle, visual style, interaction approach, and technical route were entirely left to the model. It was required to find the most suitable form of expression for the subject rather than mechanically execute a checklist of requirements.
4. Delivery standard: The model needed to independently search for reference material, source assets, and write code, including 3D rendering logic where appropriate. It first had to assess the technical boundaries of the current runtime environment, then proceed autonomously until delivering a runnable, experiential final product. Stopping at a planning document or mockup stage was not allowed.

The core of this prompt engineering approach was delegation. By handing over narrative, visual, interaction, and technology stack decisions entirely to the model, we aimed to maximize the test of its autonomous planning and engineering delivery capabilities under complex constraints.

2. Engineering Breakdown of 3D Rendering and Interactive Narrative

After receiving the instruction, the model produced two distinctly different engineering and narrative routes. Both approaches used publicly available real 3D scan data of the David sculpture as the underlying asset, ensuring visual accuracy and historical fidelity.

Approach A: Immersive Dark-Room Interaction — “The Remaining Form”

This approach took a highly artistic and philosophical direction. The model distilled the core narrative into a single idea: sculpture is what remains in the stone.

• Visual and interaction design: The page was designed as a dark room with a raw marble block standing at the center of the screen. Users press and hold a virtual button to chisel. As the interaction progresses, the stone falls away in particle-based fragments, and the statue gradually emerges from the shadows.
• Interaction detail: The model included an Undo button in the interface. When clicked, the system displays a message: “Marble has no undo.” This design choice uses technical means to embody the irreversibility of sculpting, significantly enhancing the emotional impact of the work.

Approach B: Exhibition-Style Interactive Reading — “The Man Within the Stone”

This approach leaned toward a serious digital exhibition format, emphasizing structured knowledge delivery.

• Visual and interaction design: It used a paper-like background and divided the sculpting process into four acts. By dragging a slider on the page, users can observe how the 3D stone model transforms layer by layer from a complete block into the final David statue through progressive subtraction.
• Technical implementation: This approach focused on progressive 3D rendering, using mesh slicing of STL scan data combined with a timeline-based animation for smooth transitions.

From a technical standpoint, both approaches required the model to process 3D mesh data, integrate Web3D graphics libraries such as a general WebGL framework, and implement complex interactive event flows. The model not only generated front-end page code, but also organized static assets and rendering logic appropriately, completing the full loop from raw data to visualization.

3. Cross-Modal Extension: From 3D Website to Short Video Generation

To further test the model’s coherence in cross-modal content generation, we added a video generation task after the website was delivered.

The prompt required the model to create a 60-to-90-second Chinese short video in 9:16 vertical format based on the completed website and its research data. The task was to communicate the essence of the website to a general audience and spark interest in sculpture. The model was allowed to reuse 3D assets and code from the website, and could independently decide the narrative shots, pacing, and audiovisual approach. The final deliverables were a playable MP4 file and an editable project.

At this stage, the model demonstrated strong context retention and asset reuse capabilities. It extracted the core 3D scenes from the website as video frames, reorganized the narrative, and generated a corresponding video composition script. Although final output quality still has room for improvement due to current compute quota limits and restricted access to some advanced video generation APIs, the automated pipeline of understanding requirements, extracting assets, restructuring the narrative, and generating a video project has been validated end to end.

4. Practical Findings and Implications for Enterprise Development

Through this full-pipeline test, we identified several key insights regarding the use of large language models in real-world development.

1. Balancing Cost and Latency Remains a Core Engineering Challenge

In complex 3D generation tasks, the reasoning capability of the underlying model directly determines code quality and creative interaction design. However, top-tier reasoning models often come with high compute costs and noticeable response latency. The test showed that the new-generation model has made significant progress in performance optimization: while maintaining the same or even higher logical reasoning capability, its inference cost dropped by roughly half, and response speed improved substantially. For enterprise applications, this means that when building AI agents for long-chain development, teams can adopt tiered task scheduling. Lower-cost models can be used during the ideation phase, while high-precision models are invoked for critical code review, enabling more precise control of overall R&D costs.

2. From Idea Generation to Hard Delivery

Early large models were mostly used to generate code snippets or solution prototypes, often exhibiting a gap between ideas and independent delivery. This test validated the closed-loop capability of modern intelligent programming tools: from understanding an abstract philosophical proposition such as the essence of carving, to retrieving real open-source 3D assets, to writing runnable Web3D code and handling edge cases such as the design of the Undo prompt, the model demonstrated full-stack delivery awareness. Developers can now devote more energy to requirement definition and architectural boundary control while delegating the tedious implementation details to AI.

3. Enterprise-Grade Cloud Infrastructure Is Essential

When an AI agent autonomously builds 3D websites and generates videos, it inevitably involves reading, writing, and hosting large volumes of unstructured data such as 3D scan files, texture maps, and video project files. In a real enterprise R&D environment, the stable operation of such AI agents depends heavily on an underlying cloud-native architecture. For example, highly available object storage can host massive 3D assets and audiovisual materials, container platforms can enable dynamic scaling of isolated AI code sandboxes, and logging and monitoring services can track model inference and deployment status in real time. A robust cloud infrastructure is the foundation for moving large model output from experimental generation to production-grade delivery.

In summary, the engineering capabilities of large language models in 3D interaction and multimodal content generation are maturing rapidly. With appropriate prompt constraints and cloud compute orchestration, developers can use AI to close the loop from creative ideation to final product delivery. This opens up a new R&D paradigm for digital content creation, interactive exhibitions, and front-end engineering.