Evaluating LLM 3D Creation Capabilities: From Voxel Modeling to Interactive Web Delivery

    This article explores the shift from code generation evaluation to assessing Large Language Models' 3D creation capabilities. By testing a next-generation LLM's ability to generate a voxel-style 3D building and an interactive web deliverable, the article outlines a comprehensive evaluation framework. The test highlights the model's strengths in spatial reasoning and autonomous interactive feature design, while also addressing limitations in text rendering and API quota constraints.

    1. Testing Background: A New Evaluation Direction Beyond Code Generation

    For some time, the evaluation of Large Language Models (LLMs) has predominantly focused on code generation capabilities—such as function completion, bug fixing, and algorithm implementation. However, there is a noticeable industry shift toward 3D content creation, including 3D modeling, scene rendering, and spatial reasoning, which rigorously test a model’s “spatial imagination.”

    This transition is logical. The gap in code generation among mainstream models is narrowing, whereas 3D creation demands entirely different capabilities. It requires the model to understand geometric structures and possess spatial layout reasoning—such as where stairs should lead, the vertical relationship between a ground-floor shop and a second-floor office, and whether the entrance orientation aligns with architectural common sense. These are the true litmus tests for the combination of multimodal and reasoning capabilities.

    The subject of this test is a next-generation LLM recently released on a platform. Its official materials highlight 3D creation and spatial reasoning as key capability directions, offering a dedicated 3D generation entry point. This entry point supports three forms of deliverables: modeling engineering files, interactive web pages, and printable files. To reflect the model’s native capabilities, this test utilized raw prompts without mounting any external skill plugins or pre-set templates.

    2. Prompt Design and Task Setup

    The quality of prompt design directly dictates the upper limit of the generation results. The prompt for this test was carefully designed, incorporating the following key elements:

    • Theme Anchoring: Inspired by a detective agency building from a classic detective animation. Such reference objects are highly model-friendly because they feature clear architectural structures (a ground-floor cafe and a second-floor agency), distinct visual characteristics, and ample knowledge reserves the model can draw upon.
    • Style Constraint: Explicitly required the use of a voxel block style. The voxel style is highly amenable to AI generation because it discretizes continuous geometric problems into block stacking problems, significantly reducing modeling difficulty while ensuring visual consistency.
    • Deliverable Specifications: Explicitly demanded editable 3D engineering files and rendered images, rather than just a single visual. “Editable” is crucial because only with the engineering files can users modify materials and adjust structures in subsequent workflows.

    3. Delivery Result Analysis: Fidelity and Completeness

    The model ultimately delivered two rendered images and an editable 3D engineering file. In terms of fidelity, the results were quite impressive:

    1. Accurate Architectural Structure: The classic layout from the original work was fully reproduced—the independent side staircase leading to the second floor, the detective agency on the second floor, and the cafe on the first floor. The spatial vertical relationships were entirely correct, indicating no obvious spatial reasoning errors by the model.
    2. Good Details: Two character standees were even placed at the entrance, corresponding to the main characters in the original work. Although one character’s hairstyle was treated quite abstractly (requiring annotations to recognize upon first glance), it remains within acceptable bounds as a simplified voxel-style character expression.
    3. Noticeable Omission: The model failed to generate the most iconic element of the scene—the signboard text on the agency’s wall. The signboard is the most recognizable part of the building, and its absence makes the work feel “similar but not quite.”

    This omission is quite representative: AI models perform excellently when handling structural geometry but tend to struggle with text-based decorative elements. This aligns with the consistent difficulties of text rendering in multimodal generation. Therefore, while this test cannot score a perfect mark, the overall completeness remains commendable.

    4. Beyond Expectations: Interactive Web Delivery

    What was truly surprising was the result after introducing the supplementary requirement to “generate an interactive page that can be opened in a browser.” The prompt only mentioned “opening the model file in a web page” without any functional specifications, yet the model autonomously inferred and implemented a full suite of interactive features:

    • Basic Interaction: Model drag-to-rotate and auto-rotate display, which are standard for 3D previews.
    • Scene Annotation: Added clickable text labels for key elements in the scene (the agency, cafe, characters, etc.).
    • Perspective Focusing: A sidebar listed three core elements; clicking them triggered the camera to automatically fly and focus on the corresponding object.
    • Resource Aggregation: The page integrated a model download entry and toggle displays for the two rendered images.
    • Onboarding Guide: A drag operation guide appeared upon first opening the page, lowering the usage barrier.

    Among these, the scene annotation and perspective focusing features are particularly noteworthy. In smart visualization projects—such as digital twins, campus 3D modeling, and smart city dashboards—users need to quickly locate points of interest within complex 3D scenes. Annotation plus focusing is a rigid demand for such scenarios. The model’s ability to deduce these “industry-conventional” functional requirements without explicit instructions suggests that its training data contains a considerable scale of visualization project paradigms and exhibits a degree of product thinking.

    Also worth mentioning is the visual quality of the rendered images: the filter tones and color grading were handled well, and the lighting and shadow atmosphere were spot on, making them directly usable as presentation materials.

    5. Methodology Summary: Systematically Evaluating 3D Creation Capabilities

    Based on this test, a general framework for evaluating LLM 3D generation capabilities can be distilled:

    Evaluation Dimension Key Inspection Points
    Spatial Reasoning Whether the building’s vertical/horizontal layout aligns with common sense and the reference object
    Structural Completeness Whether key iconic elements are present; any obvious omissions
    Engineering Usability Whether the delivered files are editable and available for secondary processing
    Interactive Value-add Whether the web end provides practical functions beyond basic previewing
    Visual Quality Whether the rendered lighting, shadows, filters, and composition meet presentation standards

    Additionally, there are a few takeaways for prompt writing: explicitly specify the art style (e.g., voxel) to lower generation difficulty; clearly list the deliverables checklist rather than vaguely saying “make a model”; and select themes for which the model has ample knowledge reserves to avoid overly niche reference objects.

    6. Remaining Issues and Future Observations

    The test also exposed some practical constraints. During the early release of this model, many users reported tight API quotas. The platform used for this test also had daily quota limits, forcing the postponement of originally planned tests for larger-scale scenes like real estate sand tables and smart cities. This reminds us that while evaluating model capabilities, service availability and quota strategies are equally critical factors affecting actual deployment. No matter how strong the generation capability is, if quotas cannot support batch production, the engineering value is significantly diminished.

    Overall, this test demonstrated substantive progress: moving from a single model file to a complete delivery chain of “engineering file + rendered image + interactive web page” signifies that LLMs are evolving from “generating content” to “generating deliverable products.” For teams needing to rapidly prototype 3D visualizations, such capabilities now possess the potential to integrate into actual workflows.