Evaluating Large Language Models for 3D Creation and Spatial Reasoning

    A practical framework for evaluating large language models on 3D creation, spatial understanding, editable project delivery, rendering quality, interactive web packaging, and resource efficiency. The article uses a voxel detective-office diorama as an example to show why semantic completeness and engineering usability matter as much as visual quality.

    1. Background: 3D Creation as a New Lens for Model Capability

    After a new version of a large language model was released overnight, community discussion quickly expanded from “How well does it write code?” to “Can it create 3D content?”

    This shift is not accidental. As code generation, text summarization, and routine question answering become baseline capabilities, evaluation naturally moves toward more complex areas such as spatial reasoning, cross-modal delivery, and engineering editability. 3D creation covers semantic understanding, spatial layout, asset organization, rendering, and interaction packaging at the same time, making it a high-information test surface.

    When public materials position 3D creation as a key capability, developers should maintain an engineering perspective. The question is not whether a single rendered image looks impressive, but whether the deliverables are complete, the semantics are accurate, the project remains editable, the web interaction is usable, and resource consumption is manageable. This article does not recommend any specific commercial product. The process described here can be applied to evaluating similar models and platforms.

    2. Test Design: A Bare Prompt and Three Types of Deliverables

    The original test used a “bare prompt” approach. No additional skills were attached, no reference image was provided, and no complex workflow was injected. The model received only the theme, visual style, and delivery requirements.

    The prompt essentially asked the model to use 3D modeling software to create a voxel-style miniature diorama of a detective office, make the office the visual focus, express the scene through voxel blocks, and deliver both an editable 3D project and rendered images.

    This design makes it possible to observe the model’s default understanding rather than a result produced through extensive manual constraints. The tester then added a further requirement: place the output in an HTML page that could be opened in a browser.

    The final deliverables included two rendered images, one editable 3D project file, and one interactive web page. The key issue was not simply which formats were generated, but whether the model could connect the theme, space, assets, and deliverables into a complete workflow.

    3. Deliverable Analysis: Semantic Completeness Before Visual Appeal

    The model’s result included many common elements associated with a detective office. It had a staircase along the side leading to the second floor, a detective office on the second floor, a coffee shop on the first floor, and character figures near the entrance. This indicates that the model had some ability to decompose a composite theme spatially, translating it into floors, entrances, stairs, commercial space, and character details.

    However, the main weakness was also clear: the prominent wall sign associated with the fictional setting was missing. The sign was not merely decorative. It was a core semantic anchor for the scene. Without it, the building still looked like “a detective-office-themed structure,” but not like “that particular detective office.”

    This exposes a common issue in bare-prompt evaluation: models often prioritize making the overall scene look appropriate while omitting the most recognizable semantic symbol.

    For this reason, evaluation should not stop at a first-glance review of the rendered images. A more appropriate acceptance order is:

    1. Semantic completeness: Are the core signs, entrance, floor functions, and character relationships present?
    2. Spatial consistency: Does the staircase connect the correct floors, and are the detective office and coffee shop separated logically?
    3. Engineering editability: Are objects named clearly, and can the hierarchy, materials, and cameras be adjusted later?
    4. Visual quality: Do composition, lighting, materials, and filters support the theme?
    5. Interaction usability: Can users open, drag, rotate, focus on, and download assets from the web page?

    Scoring only the rendered images can hide the usability of the underlying project. Conversely, reviewing only the project file may overlook the quality of the final presentation. A 3D generation evaluation should score semantics, spatial relationships, engineering quality, presentation, and interaction separately rather than assigning one vague overall score.

    4. The Additional Value of an Interactive Web Page

    The tester only required the files to be placed in an HTML page that could be opened. However, the model also generated features such as drag interaction, automatic rotation, scene annotations, viewpoint focusing, model downloads, rendered-image switching, and first-open guidance.

    These features are relevant to scenarios such as smart visualization, real estate models, and city modeling. The generated result becomes more than a static asset: it becomes a lightweight, presentable visualization layer that supports annotations and focused viewing.

    For example, scene annotations can associate objects such as “second-floor detective office,” “first-floor coffee shop,” and “entrance characters” with their positions in the model. Viewpoint focusing can turn a complex model into a sequence of explainable views. Model downloads and image switching can also simplify review, delivery, and secondary processing.

    These capabilities should be assessed through interaction usability. The key questions are whether annotations are accurate, whether focus actions correspond to the intended objects, whether downloaded assets are complete, and whether the initial instructions are clear—not simply whether the page looks visually elaborate.

    5. A Reusable Evaluation Framework: Six Dimensions and an Acceptance Checklist

    To make 3D generation evaluation more reproducible, teams can define structured tasks and acceptance criteria. An example configuration is shown below:

    1. {
    2. "task": "voxel_diorama",
    3. "theme": "detective_office",
    4. "required_elements": [
    5. "second-floor detective office",
    6. "first-floor coffee shop",
    7. "side staircase",
    8. "entrance characters",
    9. "wall sign"
    10. ],
    11. "deliverables": [
    12. "editable_3d_project",
    13. "render_images",
    14. "interactive_html"
    15. ],
    16. "acceptance": {
    17. "semantic_completeness": "All core elements are present and positioned appropriately",
    18. "spatial_consistency": "Floor, staircase, and entrance relationships are correct",
    19. "engineering_usability": "The object hierarchy is clear and the project remains editable",
    20. "render_quality": "Lighting, materials, and composition are stable",
    21. "interaction_usability": "Dragging, rotation, focus, and downloads work correctly",
    22. "resource_cost": "Generation time and usage quota remain manageable"
    23. }
    24. }

    The corresponding six scoring dimensions are:

    • Semantic completeness
    • Spatial consistency
    • Engineering usability
    • Rendering quality
    • Interaction usability
    • Resource cost

    Each dimension can be further divided into checkable items. This significantly reduces the subjectivity of evaluation based only on personal impressions.

    6. Engineering Practice: Structured Prompts, Asset Management, and Quota Governance

    First, prompts should explicitly list elements that must not be omitted. For example, “wall sign” can be included in the required-element checklist, with a note that the wording may be abstract but the sign itself must remain recognizable.

    Second, deliverables should be reviewed in stages:

    1. Review the engineering structure.
    2. Review the rendered images.
    3. Review the interactive web page.

    Third, the interactive page should be designed to open offline or be deployed independently whenever possible, avoiding uncontrolled dependencies on external resources.

    Fourth, asset naming and directory structures should be standardized so that materials, camera angles, and secondary renders can be adjusted or replaced later.

    If this type of evaluation is productized, general cloud capabilities such as object storage, container platforms, logging, monitoring, and alerting can be used on platforms such as Baidu AI Cloud to host generated assets, rendering tasks, and runtime observability. Specific service choices should still be evaluated according to the team’s cost, compliance, and operations requirements.

    Resource usage also requires attention. 3D generation and web packaging typically consume more resources than text-only tasks. A recommended governance process is to manage work through task queues, failure retries, result archiving, and quota alerts rather than submitting a large batch of tasks at once.

    7. Conclusion

    The value of this bare-prompt test is not simply whether a model can produce an attractive 3D model. It reveals several priorities for evaluating 3D generation:

    • Semantic completeness is more important than visual impact alone.
    • An editable project has greater long-term value than a single rendered image.
    • An interactive web page can turn a generated asset into a presentable deliverable.

    The model’s ability to add interaction features that were not explicitly required suggests that its awareness of engineering packaging is improving. At the same time, the missing core sign shows that spatial and semantic alignment remains a critical acceptance criterion.

    Structured prompts, staged acceptance, and quota governance can turn an exploratory 3D generation test into a reusable evaluation methodology.