Evaluating LLM Capabilities for 3D Creation and Interactive Web Delivery

    This article examines how LLM evaluation is shifting from single-point generation tasks to full deliverable sets, using a voxel-style miniature scene as an example. It covers how to design prompts as acceptance contracts, how to verify 3D models for structure, visual anchors, voxel language consistency, and editable project files, how to evaluate interactive HTML pages, and how to apply structured checklists and cost controls to make evaluation reproducible and engineering-ready.

    1. Evaluation Targets Are Shifting from Single-Point Generation to Deliverable Sets

    A notable shift has emerged in how large model capabilities are evaluated. Earlier benchmarks focused heavily on text and code tasks such as code completion, function generation, and unit test pass rates. As multimodal and spatial reasoning capabilities mature, 3D models, editable project files, rendered images, and standalone interactive web pages are becoming evaluation targets that better reflect real project delivery. The reason is straightforward: in real development scenarios, users rarely want code that merely “looks correct.” They want a set of deliverables that can be opened, inspected, modified, reused, and demonstrated.

    Take a voxel-style miniature scene as an example. A reasonably complete deliverable set typically includes:

    1. An editable 3D project file;
    2. At least two rendered images from different angles;
    3. A standalone interactive HTML page;
    4. Necessary model documentation, object naming, and resource organization;
    5. Reproducible prompts, versions, and generation records.

    Evaluation therefore cannot stop at “did it generate a model.” It should be broken into multiple dimensions: structural fidelity, completeness of key visual anchors, consistency of the voxel language, project editability, web interaction usability, resource download and switching capability, plus cost and reproducibility. Only by verifying each dimension separately can you avoid being misled by a single polished render or one flashy interaction that hides critical defects.

    2. Prompt Design: Upgrading from Topic Description to Acceptance Contract

    In earlier usage patterns, users often provided a bare prompt such as “use a certain 3D modeling tool to create a voxel miniature scene on a given topic, and deliver the editable project and renders.” This kind of prompt tests the model’s default understanding, but it easily misses key elements. Common issues in practice: the main building and functional zoning are largely present, but visual anchors such as the theme signboard, entrance figures, or signage text are missing; or the model file exists while the interactive web page was never explicitly requested, leaving the deliverable set incomplete.

    A more reliable approach is to write the prompt as an acceptance contract rather than a plain topic description. An example template:

    1. Task: Generate a <topic> voxel-style miniature scene.
    2. Deliverables:
    3. - Editable 3D project file;
    4. - At least two rendered images from different angles;
    5. - One standalone interactive HTML page.
    6. Key visual anchors:
    7. - The theme signboard must be visible and reasonably positioned;
    8. - The main building's floor count, side staircase, ground-floor commercial space, and second-floor office space must be identifiable;
    9. - Entrance figure decorations should be preserved, consistent with the voxel style.
    10. Interactive page requirements:
    11. - Support drag-to-rotate and auto-rotate;
    12. - Support scene annotations and view focusing;
    13. - Support model download and render switching;
    14. - Provide a drag hint on first open.
    15. Acceptance criteria:
    16. - Missing key anchors is a hard failure;
    17. - The project file must have a clear object hierarchy and adjustable material parameters;
    18. - The page must open offline, with no resources depending on inaccessible external links.

    The value of this template is that it turns “what I expect to see” into “what the model must deliver” up front. Bare prompts are suitable for probing the model’s default capability boundaries; structured prompts are suitable for engineering acceptance. The two can be combined: first use a bare prompt to observe baseline understanding, then use supplementary prompts to check how well the model follows the delivery contract.

    3. 3D Model Acceptance: Structure, Anchors, Voxel Language, and Editable Projects

    Verification of a 3D model should start with structure before details. At the structural level, check whether the main building’s floor count, side staircase, functional zoning, entrance placement, and figure decorations are all present. For miniature scenes, visual anchors are especially important: the theme signboard, signage, entrance, and staircase often determine whether the work “looks like the topic.” If the building silhouette is acceptable but the most important signboard is missing, it should not receive full marks; that is a fidelity problem, not a style problem.

    The voxel language also needs to be consistent. A voxel block style requires forms built from relatively uniform block units, with materials and lighting that serve the low-poly, blocky expression rather than mixing in excessive realistic materials. For renders, check composition, angles, filters, and readability: do the two images cover different viewpoints, is the subject clear, does the background distract, and is the signboard legible?

    The editable project file is an easily overlooked acceptance item. A model that “can be viewed” is not equivalent to a project that “can be modified.” Checkpoints include whether the object hierarchy is grouped by building, staircase, figures, and signboard; whether material parameters are adjustable; whether cameras and lights are preserved; and whether there are inaccessible external dependencies. If the project opens with chaotic object naming and baked, non-editable materials, the cost of subsequent secondary development rises significantly.

    4. Interactive Web Page Acceptance: Basic Interaction, Scene Annotations, and Resource Management

    The interactive web page is often a bonus item, but it should still be part of acceptance. Basic interactions include drag-to-rotate, auto-rotate, zoom, and a first-open guide. Scene annotations are valuable for general scenarios such as smart visualization, real estate sandboxes, and city modeling: annotations link key model elements to explanatory information, reducing demonstration cost. Sidebar view focusing further improves usability, since users no longer need to hunt for target objects manually.

    Resource management matters equally. The page should support model download and render switching, and provide graceful fallback messages when resources fail to load. If the page adds annotations, focusing, downloads, and image switching on its own when the prompt did not explicitly require them, that indicates the model has a good default understanding of “deliverable web pages,” which is a positive signal. However, you should still check whether these features are maintainable—for example, whether annotation data is centrally managed, whether download links use relative paths, and whether the switching logic is easy to extend.

    5. Structured Acceptance and Cost Control Recommendations

    To reduce subjective judgment, you can build an acceptance checklist in JSON form:

    1. {
    2. "prompt_version": "v1",
    3. "deliverables": ["editable_3d_project", "render_png", "interactive_html"],
    4. "visual_anchors": ["theme_signboard", "second_floor_office", "ground_floor_commercial", "side_staircase", "entrance_figures"],
    5. "interaction_checks": ["drag_rotate", "auto_rotate", "scene_annotations", "view_focus", "model_download", "image_switch", "first_open_guide"],
    6. "editable_checks": ["object_hierarchy", "material_params", "camera_lighting", "no_external_dependency"]
    7. }

    Automation can check whether files exist, whether the HTML opens, whether object names contain key anchors, and whether download links work; humans then judge fidelity and visual quality. When quotas are limited, evaluate in batches: test small scenes and key anchors first, then complex scenes and web interactions. Record each prompt, version, consumption, and output to avoid a single long session that leaves coverage gaps. Only by moving from “can generate” to “verifiable, reproducible, and maintainable” does 3D creation capability evaluation truly align with engineering practice.