Evaluating LLMs for 3D Creation and Interactive Web Generation: A Practical Bare-Prompt Test

    This article presents a bare-prompt experiment in which a large model is asked to create a detective-agency-themed voxel diorama and deliver an editable 3D project, rendered images, and an interactive web page. It examines the engineering challenges of 3D generation evaluation, proposes separate acceptance criteria for models, renders, and web deliverables, and distills a reproducible workflow based on prompt templates, layered checklists, and recorded evidence.

    1. Background: 3D Creation Is Becoming a New Entry Point for LLM Evaluation

    For some time, large model evaluations have focused mainly on code generation, text understanding, structured output, and similar areas. As multimodal and spatial reasoning capabilities develop, more developers are using 3D modeling, voxel scenes, editable project files, and interactive web pages to observe a model’s overall planning ability. These tasks involve requirement understanding, spatial layout, resource organization, frontend implementation, and deliverable management at the same time. They are therefore closer to real engineering scenarios than a single text or code task.

    This article discusses a “bare prompt” experiment: no additional skills are attached, no template is provided, and the goal is described only in natural language. The model under test is asked to generate a voxel diorama themed around a “detective agency” and to deliver an editable 3D project, rendered images, and an openable interactive web page. The focus of the experiment is not to evaluate a specific product, but to extract a reusable evaluation and acceptance method.

    From an engineering perspective, 3D creation tasks have several natural difficulties. First, the deliverables are diverse, including both model files and images and web pages. Second, aesthetics and structural completeness are subjective and need to be broken down into checkable items. Third, engineering details such as editability, resource paths, and interactive functions can easily be hidden by renderings that “look good.” Fourth, usage quotas and generation time can affect test coverage. Therefore, evaluation should not rely only on the final screenshot; it should establish layered acceptance criteria.

    2. Experiment Design: Observing Default Planning Ability with a Bare Prompt

    The value of a bare prompt is that it reduces interference from external scaffolding and shows whether the model proactively fills in reasonable requirements. In the experiment, the prompt only asked for the detective agency theme to be presented in a voxel block language, for the building’s visual highlights to be emphasized, and for an editable 3D project and rendered images to be delivered. A follow-up requirement then asked for an interactive display page that could be opened in a web browser.

    To improve reproducibility, it is advisable to split the prompt into five layers: goal, constraints, deliverables, acceptance criteria, and interaction requirements. For example:

    1. Goal: Create a voxel diorama themed around a "detective agency."
    2. Visual constraints: two-story agency, ground-floor coffee shop, side staircase, character at the entrance, using a voxel block language.
    3. Key anchor: the building's exterior wall must have a clearly visible "detective agency" sign.
    4. Deliverables: editable 3D project file, 2 rendered images, 1 interactive web page.
    5. Interaction requirements: drag-to-rotate, auto-rotate, scene annotations, view focus, model download, image switching.
    6. Acceptance: files can be opened; scene elements are complete; the page runs without additional configuration.

    This template is not complex, but it makes the “key visual anchors” and “page functions” explicit. A typical problem in the original experiment was that the model recreated the two-story agency, the ground-floor coffee shop, the side staircase, and even the character at the entrance, but omitted the most prominent sign on the wall. This omission suggests that the model may be better at generating structures based on common architectural semantics, while a user’s “visual anchor” needs to be explicitly emphasized. During evaluation, “whether key elements appear” should be an independent scoring item rather than being folded into a general impression of overall fidelity.

    3. Deliverable Acceptance: Models, Renders, and Interactive Web Pages Should Be Assessed Separately

    3.1 3D Model: Structural Completeness and Editability Matter Equally

    When accepting a 3D model, the first step is to check whether the file can be opened normally in general-purpose 3D modeling software, whether the scene hierarchy is clear, and whether the model is composed of editable meshes or voxel units rather than a single indivisible object. Next, check the building structure: whether the two-story agency, ground-floor coffee shop, side staircase, entrance character, and other elements are all present. Then check key anchors, such as whether signs, doors, windows, and markers appear. Finally, check whether coordinates, scale, materials, and lighting are reasonable, and whether there is obvious clipping, floating geometry, or abnormal normals.

    Rendered images can create the illusion of high completion, but they cannot replace an editability check. Even if a model renders beautifully, if the project file has a messy hierarchy, missing materials, or incorrect resource paths, the cost of later editing remains high. It is therefore recommended to score the “editable project file” and the “rendered images” as two independent deliverables.

    3.2 Rendered Images: Focus on Composition, Lighting, and Consistency

    Rendered images can be accepted from three perspectives: whether the composition highlights the main subject, whether the lighting makes the voxel structure clearly legible, and whether multiple images maintain scene consistency. The rendered images in the experiment had good filter effects, indicating that the model has some planning ability for post-processing stylization. However, evaluation should avoid scoring solely by visual impression. Observable indicators such as subject proportion, shadow readability, material distinction, noise, and overexposure can be recorded.

    3.3 Interactive Web Page: Unrequested Functions Are Both a Bonus and a Risk

    In the experiment, the web page exceeded the minimum requirements. In addition to model dragging and auto-rotation, it provided scene annotations, view focus, model download, image switching, and first-open guidance. These functions have practical reference value for visualization scenarios such as real-estate sandbox models, city modeling, and equipment display. Scene annotations and view focus can upgrade “model display” into an “explainable 3D dashboard,” while download and image switching make delivery and archiving more convenient.

    However, the more functions there are, the more acceptance should focus on maintainability and security. For example, check whether resources are loaded locally, whether the page is usable on mobile, whether interaction prompts can be closed, whether annotations obscure the main subject, and whether download links are valid. For enterprise applications, it is also necessary to check whether the page introduces unnecessary external dependencies, whether it has basic error messages, and whether there is a loading state when model resources are large.

    An acceptance checklist can be designed as follows:

    1. {
    2. "files": ["editable 3D project", "rendered images", "web page"],
    3. "model": ["opens", "clear hierarchy", "all key anchors present", "no clipping"],
    4. "render": ["subject highlighted", "reasonable lighting", "consistent across images"],
    5. "web": ["drag-to-rotate", "auto-rotate", "scene annotations", "view focus", "download", "image switching", "mobile usable"],
    6. "compliance": ["no unauthorized IP elements", "no suspicious external dependencies", "portable resource paths"]
    7. }

    4. Extracting a Reproducible Evaluation Method from One Experience

    First, fix the variables. Record the model version, full prompt, generation time, deliverable list, and page screenshots, rather than keeping only an impression. Second, score in layers. The total score can be divided into dimensions such as requirement satisfaction, structural completeness, editability, render quality, interactive functions, maintainability, and compliance, with observable evidence for each item. Third, make key constraints explicit. Visual anchors such as brands, signs, characters, proportions, and color schemes should be listed separately, and positive and negative examples can be provided when necessary. Fourth, manage usage quotas. 3D generation is usually more time-consuming and resource-intensive than plain text. Before testing, scenarios should be prioritized, core deliverables should be verified first, and complex interactions should be expanded afterward.

    In addition, when copyrighted themes, characters, or logos are involved, authorization should be obtained or generic elements should be used instead. Evaluation articles and examples should avoid retaining recognizable IP names, so that a technical test does not become an infringement risk.

    5. Conclusion

    This experiment shows that large models already have a certain degree of engineering usability in 3D scene planning, voxel-style expression, and web interaction completion. They can even proactively add functions such as scene annotations, view focus, download, and image switching. However, key visual anchors still need to be explicitly constrained in the prompt, and deliverable acceptance cannot rely only on rendered images. Only by turning one experience into a “prompt template + layered acceptance checklist + reproducible record” can 3D creation ability be evaluated more objectively and a reusable engineering method be provided for subsequent visualization projects.