Testing LLM 3D Content Generation: Two AI Models Recreate the Making of Michelangelo's David

From a Block of Stone to a Statue: An Ambitious Digital Creation with LLMs

As 3D content generation and agent technology evolve at a rapid pace, how do you evaluate what a large language model can actually do across the full pipeline from creative idea to deliverable work? Recently, I ran an in-depth experiment around the theme “the birth of Michelangelo’s David”: the AI had to understand the historical context, then autonomously design the narrative, visual style, and technical implementation of an interactive 3D website, and finally ship a working product. This article walks through the full experiment—the process, the results, and the pitfalls—and explores where today’s LLMs stand when it comes to complex creative tasks.

1. The Creative Brief: Making AI Understand the Essence of Carving

The core instruction was simple but highly open-ended: build an interactive Chinese-language website that helps ordinary people understand how the David emerged from a block of stone. The key requirement was that viewers should not just memorize a few historical dates—they should come away with a new understanding of carving itself: that sculpture is the art of subtraction, in which the artist “releases” a human form from a chaotic mass of rock.

To give the model full creative freedom, the brief explicitly stated that the narrative angle, visual style, interaction design, and technical approach were entirely up to the AI, and that the deliverable had to be a runnable, experiential web page—not a proposal or a mockup. This effectively tested three core capabilities: deep understanding of the subject, frontend and 3D engineering skills, and product decision-making in the face of ambiguous requirements.

2. Two Models, Two Solutions: A Subtractive Narrative vs. an Exhibition-Style Deconstruction

Two different models, working independently, produced two sharply different design philosophies—mirroring the two classic content paradigms of “immersive experience” and “rational explanation.”

Solution 1: “Remnant Form” in a dark room

The first model’s piece, titled Remnant Form, framed the website as a dark room with an unformed block of marble suspended at its center. Users press and hold the mouse button to simulate a chisel strike; with each press, the stone peels away in particles or fragments, and David’s body gradually emerges.

The brilliance of this design lies in its copywriting. The page offers an “Undo” button, but when clicked, nothing is undone. Instead, a cold message pops up: “Marble has no undo.” The detail sets the irreversibility of the physical world against the reversibility of the digital one, instantly giving the work philosophical depth—it is no longer a simple history lesson, but an invitation to reflect on the cost and inevitability of creation itself.

Solution 2: “Figure in the Stone,” a four-act drama

The second model took a more traditional but tightly structured “exhibition reading” route, titled Figure in the Stone. The page uses a beige, paper-like background to create the atmosphere of reading archival documents, and divides the sculpting process into four acts: quarrying, rough carving, fine carving, and completion. By dragging a vertical slider, users can watch the 3D model transform from a squared rough block into David’s precise silhouette, layer by layer.

This path leans toward “understanding.” The reversible interaction—dragging the slider—lets viewers repeatedly observe the stone diminishing, and the accompanying text strikes a good balance between visual intuition and knowledge transfer.

Notably, both websites were built on the same David model data sourced from a public 3D scanning library. This kept the comparison fair, and it showed that with openly available assets, the differences between models now lie mainly in narrative ability and engineering execution.

3. Technical Implementation: Key Points and Shared Challenges

Although the two designs look nothing alike, both relied on mature browser-side 3D rendering (the familiar WebGL stack), and both followed the same development path: retrieve assets first, then write the core logic, then debug the interaction feedback.

One clear pattern emerged: LLMs are highly proficient at common frontend tasks such as loading 3D models and binding interaction events. The code was well structured, with essentially no syntax-level defects. The differences appeared at the level of physical realism. When simulating falling stone, the first model chose to stack many irregular geometry pieces and hide them one by one, rather than render a true physics-based particle system. That sacrificed some realism but drastically reduced performance overhead and development complexity—a sign of the model’s ability to make engineering trade-offs under resource constraints, which matters a great deal in real production.

4. From Website to Short Video: Testing Multimodal Output

After the websites were delivered, I added a follow-up test: each model had to produce a 60–90 second Chinese-language short video introducing the essence of its website, based on the finished site. The prompt noted that they could reuse assets and code from the website and were free to recompose the visuals.

This stage exposed the real gap in today’s multimodal capabilities. Thanks to deep memory of the page elements—the buttons, the 3D model, the copy—both models produced fairly complete video scripts covering storyboard design, narration, and pacing. But the actual rendering was inconsistent. The stronger performer could automatically capture the website’s interactions as video footage and deliver accurate voiceover. The other repeatedly failed mid-render—its compute resources were preempted by other tasks—and never delivered a playable MP4 file on time. The lesson: when relying on LLMs for video generation, you need to watch not only the model’s algorithmic capability but also the stability and scheduling of the compute resources behind it.

5. Findings and Practical Advice for Developers

Looking back at the whole experiment, several conclusions are worth a developer’s attention:

First, a model’s grasp of abstract concepts sets the ceiling of the work. The gap between the two models was not in 3D technique—it was in whether they could capture the core metaphor of subtractive carving. The model that designed the “no undo” interaction clearly understood what irreversibility means in artistic creation. If your goal is not just another showcase page, abstracting the core concept in your prompt works better than piling up technical terms.

Second, interaction design and narrative should be stitched together, not bolted on. A great 3D website is not a rotatable model with a block of explanatory text next to it. Binding the slider to the stone’s reduction, and the click to a chisel strike—this “interaction as content” thinking is what elevates the user experience.

Third, cost and efficiency remain unavoidable constraints. In this test, the newer, stronger model was indeed a significant improvement over the previous generation—roughly half the cost and twice the speed—but it still carried stability risks in high-concurrency or long-running generation tasks. For small and mid-sized teams, production environments need retry logic and checkpoint/resume mechanisms so that a single point of failure does not scrap an entire creative workflow.

Fourth, open-source assets are the invisible foundation of AI creation. The 3D scan used in this experiment came from an openly licensed public project—proof that high-quality public asset libraries can drastically lower the “from-zero cost” of AI creation. Going forward, re-creating with LLMs plus open datasets will become one of the mainstream modes of digital content production.

6. Closing Thoughts: The Next Stage of AI Creation Is “Principled Creation”

The highest value of this experiment is not showing how pretty a webpage an AI can build. It lies in revealing a trend: today’s generative AI already has the composite ability to move from understanding an abstract brief, to mobilizing code and 3D resources, to delivering a finished multimodal product. It is still unstable in many ways, but it can already think about narrative, weigh technical routes, and ship work with subjective expression—much like a genuine junior creator.

For developers, this means the workflow of the future will inevitably shift toward “humans set the principles and boundaries; AI handles divergence and execution.” And how to inject a touch of humanistic reflection into the work beyond the code—the way “Marble has no undo” did in this experiment—remains the irreplaceable moat of human creators.