Evaluating LLMs for 3D Voxel Miniature Landscape Generation: Practices and Methods

    As LLM capabilities in coding mature, evaluation metrics are shifting towards 3D content creation. This article explores a methodology for evaluating LLMs using "naked prompts" to generate 3D voxel miniature landscapes. It covers prompt design, restoration assessment focusing on key identification elements, spontaneous UI delivery, API quota management, and a reusable evaluation framework. The findings reveal that while models show impressive progress in spontaneous interactive web layers, they still struggle with critical signage and character details, making them best suited for accelerating prototyping rather than replacing manual modeling.

    Evaluating LLMs for 3D Voxel Miniature Landscape Generation: Practices and Methods

    For a long time, code completion, algorithmic problem solving, and scaffolding generation have been the core metrics for measuring LLM capabilities. However, as models’ coding abilities improve across the board, pure code evaluation struggles to differentiate them. Recently, the community’s evaluation focus has noticeably shifted toward 3D content creation: can a model produce editable modeling project files, rendered images, or even interactive 3D preview web pages using just a natural language prompt?

    The value of this type of evaluation lies in its simultaneous assessment of multiple model capabilities: semantic understanding of text requirements, spatial structure reasoning, code organization for procedural modeling, and control over the completeness of the final deliverables. Compared to “writing a sorting algorithm,” 3D scene generation tasks are far more open-ended and better expose a model’s shortcomings in complex, multi-step tasks.

    Many vendors have already begun promoting 3D creation and spatial reasoning as key capabilities, with some platforms offering output channels for modeling files, web pages, and 3D printing. For developers, the real question is: how do these capabilities perform in practice using naked prompts without any auxiliary plugins or skills attached?

    2. Designing a “Naked Prompt” Evaluation

    To truly reflect the model’s native capabilities, this test deliberately avoids attaching any plugins or preset skills, relying solely on a single natural language prompt. A good evaluation prompt needs to contain several key elements:

    • Theme and Style Constraints: Clearly define the scene theme and visual language (e.g., voxel style) to prevent the model from free-styling and causing style drift.
    • Reference Source: Provide a well-known design reference so that the restoration accuracy can be measured.
    • Deliverable Requirements: Explicitly list the expected outputs, such as editable project files, rendered images, interactive web pages, etc.

    An illustrative prompt structure looks like this:

    1. Use procedural modeling to create a 3D voxel miniature landscape themed around [a specific building].
    2. Inspired by [reference work], make the landmark building the visual highlight.
    3. Express it using voxel blocks, and deliver an editable modeling project file and rendered images.

    There is a notable lesson here: deliverables not explicitly mentioned in the prompt are rarely produced by the model, though there are exceptions. In this test, in addition to the requested model file and two rendered images, the model unexpectedly produced a feature-rich interactive web page—a deliverable not explicitly requested. This suggests that during evaluation, we should distinguish between “explicit requirements” and “the model’s spontaneous completion behavior,” as the latter is often an important signal of a model’s product-oriented thinking.

    3. Restoration Assessment: Focusing on “Key Identification Elements”

    Once the model files are obtained, the first assessment dimension is scene restoration. When evaluating, it is recommended to categorize scene elements into three layers:

    1. Structural Layer: Is the overall layout of the building correct? For example, floor relationships, staircase positions, and the correspondence between ground-floor shops and upper-floor offices.
    2. Decoration Layer: Are entrance figures, door and window details, and vegetation complete, and are their shapes reasonable?
    3. Signage Layer: Are the most recognizable landmark elements missing?

    The main lesson from this test lies in the signage layer: the model restored the building structure and even placed character models, but it missed the most prominent sign text from the reference scene. This “missing key identification element” is a common issue in 3D generation—the model can grasp spatial structures but is not sensitive enough to text-based or symbolic visual anchors. There were also obvious deviations in character modeling, such as the abstract representation of hairstyles differing significantly from the original.

    Therefore, when scoring, it is advisable to use a split-scoring approach of “Structure Score + Signage Score” to prevent an overall impression score from masking critical flaws.

    4. Interactive Webpage: Unexpected Spontaneous Delivery

    The biggest surprise in this test came from the 3D preview web page generated spontaneously by the model. The prompt only asked to “open the file in a web page,” yet the model delivered a fairly complete feature set:

    • Basic Interaction: Model drag-to-rotate, auto-rotate, and drag guidance for first-time users.
    • Scene Annotation: Adding text labels to key elements in the model to facilitate understanding of the scene composition.
    • View Focusing: Listing major scene elements in a sidebar, allowing the camera to focus on the corresponding position upon clicking.
    • Resource Aggregation: Integrating model downloads and multiple rendered image switches into a single page.

    The scene annotation and view focusing features are particularly noteworthy for developers. In smart visualization projects—such as building sandboxes, campus digital twins, and city modeling displays—annotation and focusing are among the most common interactive requirements. If a model can proactively implement such functions without being prompted, it indicates that its training data has absorbed good visualization engineering practices.

    Of course, such spontaneous deliveries still require manual verification: whether the element names in the focus list are accurate, and whether the annotation positions match the actual model positions, may still deviate.

    5. The Impact of API Quotas on Evaluation Pace

    A problem easily overlooked in practice is resource constraints. Many platforms set daily usage limits for new-generation models, and credit-based platforms are similarly restricted. This directly affects evaluation coverage: planned multi-scenario tests (single buildings, real estate sandboxes, smart city-level scenes) may have to be spread across multiple days due to exhausted quotas.

    Recommended response strategies include:

    • Priority Ranking: Test the core scenarios that best represent the model’s capability boundaries first, followed by extended scenarios.
    • Prompt Dry-Run: Before officially consuming quotas, use a low-cost method to verify whether the prompt wording is clear to avoid invalid calls.
    • Result Archiving: Fully archive the deliverables of each call to facilitate cross-day comparative analysis.

    6. Summary of Evaluation Methodology

    Based on this practice, a reusable 3D generation capability evaluation framework can be distilled:

    1. Use naked prompts to eliminate the interference of auxiliary tools and observe the model’s native capabilities.
    2. Explicitly define style, reference, and deliverables in the prompt, but do not exhaustively list details, leaving room to observe the model’s spontaneous behavior.
    3. Layer the restoration assessment scoring and separately verify key identification elements.
    4. Perform item-by-item manual verification of spontaneously generated interactive features.
    5. Plan quotas in advance to ensure the evaluation covers enough scene types.

    The overall conclusion is that current LLMs have shown considerable progress in 3D voxel scene generation and web interaction integration, especially in the spontaneous completion of visualization interaction layers, which is impressive. However, there are still obvious shortcomings in restoring key signage and detailed character modeling. For teams looking to introduce such capabilities into their visualization project pipelines, a more reasonable positioning at this stage is “accelerating prototype construction,” rather than replacing manual modeling and art review.