Evaluating LLM 3D Generation: From Voxel Landscapes to Interactive Web Pages

    As 3D content generation and spatial reasoning become crucial for evaluating Large Language Models (LLMs), this article explores a practical approach to testing these capabilities. By generating a voxel-style detective agency miniature landscape and a corresponding interactive web page using raw prompts, we analyze the model's ability to map natural language to 3D assets and deliver editable engineering files. The article introduces evaluation dimensions, a reusable prompt template, an acceptance checklist, and engineering improvement suggestions to transition LLM generation from experimental output to verifiable engineering deliverables.

    1. Background: 3D Creation as a New Metric for LLM Evaluation

    In the past, the evaluation of large model capabilities mainly focused on text understanding, code generation, and structured output. With the growing demand for 3D content production, spatial reasoning, and web interaction, an increasing number of practices are beginning to use “whether a usable 3D scene and interactive page can be generated from natural language” as a new perspective for observing the comprehensive capabilities of models. Unlike simply generating an image, 3D creation tasks simultaneously involve semantic understanding, spatial relationship reasoning, asset organization, engineering file delivery, and front-end interaction implementation, making it a typical multi-stage composite task.

    In a general practice, we used a “raw prompt” approach without attaching any external skills or dedicated workflows. We directly required the tested LLM to generate a voxel miniature landscape themed around a detective agency and deliver an editable 3D engineering file and rendered images. We subsequently added a requirement for a web page display. The purpose of this practice was not to evaluate a specific product, but to abstract a reusable observation methodology: can the LLM stably map scene elements from natural language to 3D assets, maintain the integrity of key elements, and deliver the model integrated with an interactive page?

    2. Practice Design: Shifting from Single Image to Multi-Deliverable Acceptance

    The core constraints of this task included: presenting the themed building using a voxel block language; highlighting the visual focal point of the detective agency; delivering an editable 3D project and rendered images; and, in the supplementary requirements, displaying the model through a web page. The key shift here is that the evaluation target is no longer just “whether the generated result looks similar,” but extends to “whether the deliverables are complete, editable, interactive, and acceptable for review.”

    From the results, the model delivered two rendered images and an editable 3D engineering file. The building volume was essentially present: elements such as external stairs, second-floor office space, and first-floor commercial space were reflected, and a character was placed at the entrance. This indicates that the LLM has a certain feasibility in mapping semantics to scene structures, capable of transforming “themed building + voxel style + miniature landscape” into a preliminary 3D asset combination.

    However, the problems were also obvious: the signature elements that are very prominent in animation or IP themes did not appear in the model, and the character’s design was relatively abstract. This exposes a common issue: when the prompt only provides the theme and style without explicitly listing a “checklist of key elements that must appear,” the model might prioritize satisfying the overall atmosphere and building volume while omitting details crucial for theme recognition. For commercial projects, such omissions often directly affect delivery acceptance.

    3. Interactive Web Capabilities: From Model Display to Scene Annotation

    After supplementing the web page requirements, the generated result exhibited a functional density beyond basic display: it supported model dragging, auto-rotation, scene annotation, camera focus, model download, render switching, and a drag guide on first open. For scenarios like smart visualization, real estate sandboxes, and city modeling, these capabilities hold practical reference value. Scene annotation can associate key objects in the model with explanatory information; camera focus can guide users to quickly locate buildings, characters, or functional areas; download and display switching facilitate review and reporting.

    This shows that the LLM can not only generate static 3D assets but also generate a layer of lightweight interaction logic around the assets. However, the risks are equally present: “exceeding expectations” in web functionality does not equate to “deliverable.” Without clear acceptance criteria, additional features may introduce uncontrollable complexity, such as inconsistent object naming, mismatched annotations and model hierarchies, insufficient mobile adaptation, and non-closable first-time guides. Therefore, engineering practices should incorporate interactive features into the acceptance checklist rather than just looking at whether the page “looks rich.”

    4. Suggested Evaluation Dimensions

    For such tasks, it is recommended to establish an acceptance checklist based on the following dimensions:

    Dimension Observation Point Acceptance Suggestion
    Semantic Fidelity Whether the themed building, signboard, characters, and functional spaces appear Establish a checklist of key elements and verify item by item
    Spatial Reasoning Whether the relationships between stairs, floors, entrances, commercial and office spaces are reasonable Check floor plans and vertical transportation
    Asset Integrity Whether there are missing, floating, intersecting, or proportionally abnormal elements Manual inspection combined with automated bounding box checks
    Editability Object hierarchy, naming, materials, and modularity Require object grouping and semantic naming
    Interactivity Whether dragging, rotating, annotating, focusing, and downloading are usable Conduct separate acceptance tests on desktop and mobile
    Maintainability Whether the page structure, configuration items, and resource references are clear Check resource paths and configuration separation
    Security & Compliance Whether there are external dependencies, copyright risks, or sensitive content Dependency whitelisting and content review

    5. Reusable Prompt Template

    To reduce the risk of omitting key elements, the prompt should be upgraded from a “descriptive request” to a “specification-based requirement.” An example is as follows:

    1. Task: Generate a voxel-style 3D miniature landscape, and deliver an editable project file, rendered images, and a web display page.
    2. Theme: Detective Agency.
    3. Must-have elements:
    4. 1. Main building structure;
    5. 2. Prominent theme signboard;
    6. 3. External stairs;
    7. 4. Second-floor office space;
    8. 5. First-floor commercial space;
    9. 6. Entrance character (stylized is acceptable).
    10. Spatial constraints: Clear floor relationships, logically connected stairs, recognizable entrance.
    11. Deliverables: Editable 3D project file, at least 2 rendered images, locally openable web display page.
    12. Interaction requirements: Drag to rotate, auto-rotate toggle, scene annotation, camera focus, model download, render switching, first-open guide.
    13. Acceptance criteria: Key elements present item by item; object naming readable; no console errors on the page; basic mobile operability.
    14. Prohibitions: Do not omit the signboard; do not generate uneditable merged meshes; do not introduce unknown external dependencies.

    6. Acceptance Checklist Example

    1. {
    2. "scene": "detective_office_voxel",
    3. "must_have": [
    4. "main_building",
    5. "theme_signboard",
    6. "external_stair",
    7. "second_floor_office",
    8. "first_floor_commercial_space",
    9. "entrance_characters"
    10. ],
    11. "deliverables": {
    12. "editable_3d_project": true,
    13. "renders": 2,
    14. "web_viewer": true
    15. },
    16. "interactions": [
    17. "drag_rotate",
    18. "auto_rotate_toggle",
    19. "scene_annotation",
    20. "camera_focus",
    21. "model_download",
    22. "render_switch",
    23. "first_open_guide"
    24. ],
    25. "acceptance": {
    26. "no_missing_key_element": true,
    27. "object_naming_readable": true,
    28. "no_console_error": true,
    29. "mobile_basic_usable": true
    30. }
    31. }

    7. Engineering Improvement Suggestions

    First, specify before generating. Break down the theme into key elements, spatial relationships, deliverables, and acceptance criteria to prevent the model from guessing priorities on its own. Second, use multi-round iteration. The first round generates the overall structure, the second round fills in key details like signboards, characters, and annotations, and the third round checks the consistency between the page and assets. Third, assign weights to key elements. Signboards, entrances, and floor relationships should serve as hard constraints rather than stylistic descriptions. Fourth, inspect the object hierarchy and naming of the editable project to ensure that subsequent art and development teams can take over. Fifth, the web page should retain a fallback plan, ensuring that the model and rendered images can still be displayed if interactions fail. Sixth, when API call quotas are limited, prioritize designing high-information-density test cases covering four types of scenarios: buildings, characters, annotations, and mobile.

    8. Conclusion

    From this general practice, it can be seen that LLMs already possess a certain feasibility in 3D miniature landscape generation, editable project delivery, and web interactive page integration. However, there is still a gap before reaching stable engineering delivery. The core issue is not “whether it can generate,” but “whether it can deliver completely according to specifications and pass acceptance.” Therefore, a more reasonable technical approach is to incorporate the LLM into a closed loop of “specification definition — generation — automated inspection — manual review — version iteration,” constraining the output through key element checklists, interactive acceptance tables, and editability checks. Only by transforming generation capabilities into verifiable engineering capabilities can 3D content generation better fit into real project workflows.