Evaluating LLMs on 3D Miniature Landscape Generation: From Bare Prompts to Deliverables

    As large language models move into real production, evaluation is shifting from "can it write usable code" to "can it deliver a complete, usable digital asset." This article uses 3D miniature landscape generation as a representative test case. It walks through a bare-prompt experiment (a detective agency scene with no plugins or preset workflows), breaks delivery into three layers (3D assets, rendered images, and an interactive web page), and explains why a small omission like a missing sign can degrade the whole result. It then offers a structured, acceptance-driven prompt template, discusses the automatic expansion of interactive web features, defines five evaluation dimensions, and closes with practical advice for engineering 3D generation into reliable, editable, deliverable assets.

    From Code Generation to 3D Delivery: A New Focus for Model Evaluation

    For a while, LLM evaluation centered on text generation, code completion, and structured output. But as generative AI enters real production scenarios, the focus is shifting from “can it write usable code” to “can it deliver a complete, usable digital asset.” A 3D miniature landscape is a representative task of this kind.

    A 3D miniature landscape is not just a static model. The model has to do several things at once: understand the theme, plan the space, build the assets, organize the scene, output rendered results, and in some cases generate an interactive web display. Compared with a simple coding Q&A, this kind of task puts the model’s overall engineering ability to the test, including:

    • Natural language understanding: extracting the theme, style, structure, and delivery requirements from a vague prompt;
    • Spatial reasoning: understanding building hierarchy, viewpoint relationships, object occlusion, and scale;
    • Asset generation: producing buildings, roads, characters, decorations, and other visual elements;
    • Interaction design: organizing drag, rotate, focus, and annotation behaviors in the web page;
    • Engineering packaging: assembling model files, renders, and web resources into deliverables.

    So 3D generation is no longer just a question of “can the model draw a picture.” It is a full “text-to-digital-scene” engineering problem.

    The Bare-Prompt Experiment: Observing Raw Capability Without Extra Workflows

    A typical experiment can start from a fairly concrete theme: use 3D modeling to create a miniature landscape themed around a “detective agency.” The result should include both an editable 3D project file and rendered images, and further form an interactive web page that opens in a browser.

    To observe the model’s raw ability, the experiment deliberately skips extra plugins, preset workflows, and manual compensation. Only a single natural-language prompt is provided, and the model must complete understanding, planning, generation, and delivery on its own.

    The value of this kind of “bare-prompt” experiment is that it reflects the model’s capability boundary without external assistance. In real business scenarios, users rarely hand over a complete specification from the start; they usually describe a theme and an expected effect first. So whether the model can derive a complete delivery structure from limited information is a key measure of its engineering usability.

    The task description in the experiment can be abstracted as follows:

    1. {
    2. "task": "Generate a detective agency-themed 3D voxel miniature landscape",
    3. "deliverables": ["Editable 3D project", "Rendered image", "Interactive web page"],
    4. "constraints": {
    5. "key_elements": ["Second-floor agency", "First-floor cafe area", "External staircase", "Wall signage", "Entrance character"],
    6. "style": "Voxel",
    7. "scale": "Miniature",
    8. "view": ["Overview", "Entrance", "Interior"]
    9. },
    10. "acceptance": {
    11. "correct_spatial_hierarchy": true,
    12. "key_signage_visible": true,
    13. "interaction_usable": true
    14. }
    15. }

    This looks like a single scene, but it already carries several engineering constraints. The model not only has to “draw it out” but also ensure the spatial structure is sound, the key elements are complete, the web page is interactive, and the deliverables are complete.

    Breaking Down the Delivery: Model, Render, and Web

    A complete delivery usually breaks into three layers: the 3D asset, the rendered images, and the interactive web page. Each should be evaluated on its own, not just judged by the final screenshot.

    1. 3D Assets: Structural Fidelity Determines Usability

    The 3D asset is the foundation for everything that follows. A good miniature landscape must first match the prompt’s description in spatial structure. For a detective agency theme, the model needs to express these relationships:

    • A multi-story building;
    • The second floor as the main agency space;
    • The first floor as a cafe or public reception area;
    • A staircase on the side or exterior leading to the second floor;
    • A character element at the entrance;
    • A recognizable theme sign on the wall or above the doorway.

    If the model can produce this structure, it shows some spatial understanding. But structure alone is not enough. You still need to judge whether the proportions are balanced, the occlusion is reasonable, the nodes are editable, and the materials are consistent.

    In this task, the result might include the staircase, the second-floor agency, the first-floor cafe area, and the entrance character, but with one typical problem: the most important wall sign is missing. This looks like a small detail, but it hurts theme recognition. The sign is the core semantic marker of the “detective agency” theme; once it is missing, the scene degrades from a “clearly themed” result into “an ordinary building model.”

    2. Rendered Images: Visual Quality Is the Main Entry Point for Perception

    Rendered images are how users perceive generation quality most directly. Even if the 3D asset has missing details, a high-quality render can still lift the overall impression.

    When evaluating a render, people usually look at:

    • Whether the composition is complete;
    • Whether the lighting is natural;
    • Whether the viewpoint conveys the theme;
    • Whether the voxel style is consistent;
    • Whether colors and materials are harmonious;
    • Whether the image has display value.

    In some results, the model may output renders from multiple angles by default and improve the visuals with filters, tone, and background tuning. This ability matters a lot for non-expert users, because many business stakeholders do not inspect the 3D project files directly; they judge whether a result meets expectations through the renders first.

    3. Interactive Web Page: From Static Display to Business Visualization

    If the model can also generate an interactive web page, the delivery value rises further. Such a page can typically include:

    • Model loading and display;
    • Drag-to-view;
    • Auto-rotation;
    • Zoom control;
    • Scene annotation;
    • Element focus;
    • Viewpoint switching;
    • Render switching;
    • File download entry;
    • First-use guidance.

    These look like ordinary front-end features, but they carry real weight in a 3D generation task. A 3D asset is often hard for non-technical people to understand, whereas an interactive web page can turn a complex space into an interface that can be perceived, explored, and explained.

    Diagnosing the Key Problem: Why a Missing Sign Affects the Whole Evaluation

    In 3D generation, correct structure is only the baseline; semantic completeness is what matters. A missing sign can be understood on three levels.

    1. Long-Tail Details Get Compressed by the Model

    Theme words in a prompt are usually understood first, but decorative elements, text elements, and identifying elements tend to be weakened. Models lean toward building the main structure first, such as the building, floors, staircase, and characters, and then adding details. If a detail is not explicitly constrained, the model may treat it as optional.

    A sign is not a simple geometric shape. It also involves text, textures, materials, position, and visibility. Compared with a staircase or a floor, a sign is more costly to generate and implement, and easier for the model to skip.

    2. Text-Based Assets Remain a Hard Problem

    A theme like “detective agency” relies on strongly identifying elements to be expressed, and text is exactly the area where generative models are relatively weak. The reasons include:

    • Text needs precise layout, and generative models tend to blur or distort it;
    • Text usually has to be attached to a specific surface;
    • Text content must be strongly related to the theme;
    • Text placement affects overall recognition.

    So if the prompt does not explicitly require that “the sign must be visible,” “the text must be legible,” and “the doorway must contain the theme marker,” the model will likely produce only an ordinary building and miss the text sign.

    3. The Prompt Lacks Acceptance Conditions

    Many prompts only describe “what I want to do” without defining “what counts as done.” In 3D generation, without acceptance conditions, the model fills in gaps according to its own interpretation. The result may be structurally sound but missing key business semantics.

    Prompt design should therefore be upgraded from a “theme description” to a “delivery contract.” That means stating not only the theme but also the elements that must be present, the proportions that must be met, the deliverables that must be provided, and the interactions that must be implemented.

    Prompt Optimization: From Theme Description to Acceptance-Driven

    To address missing signs, thin details, and incomplete interactions, a structured prompt can raise generation quality. An effective 3D generation prompt usually contains these modules:

    1. Task goal: what type of digital asset to generate;
    2. Theme semantics: what the core theme is;
    3. Spatial structure: how the building, floors, areas, and paths are organized;
    4. Key elements: which elements must appear;
    5. Visual style: voxel, low-poly, realistic, cartoon, and so on;
    6. Deliverables: model, renders, web page, and so on;
    7. Interaction requirements: rotate, zoom, annotate, focus, and so on;
    8. Acceptance criteria: what counts as failing if it is missing.

    It can be organized like this:

    1. Generate a 3D voxel miniature landscape themed around a detective agency.
    2. Requirements:
    3. 1. The main building has two floors: the first is a cafe reception area, the second is the agency office area.
    4. 2. The side of the building must have an external staircase leading to the second floor.
    5. 3. The entrance must have two character roles, one at the doorway and one near the staircase.
    6. 4. The exterior wall must include a prominent detective agency sign with clearly legible text.
    7. 5. The overall style is consistent voxel blocks, using a contrast of warm and cool colors.
    8. 6. Output an editable 3D project file, at least two rendered images, and one interactive web page.
    9. 7. The web page must support drag-to-view, auto-rotation, scene annotation, and focus on key areas.
    10. 8. If the sign is missing, the floor relationship is wrong, or the web page cannot be interacted with, the result is considered failing.

    Compared with a one-line request like “make a detective agency model,” the structured prompt turns vague requirements into checkable generation conditions. The model gets a clearer goal when generating, and it is easier to locate problems during evaluation.

    Automatic Expansion of the Interactive Web Page: A Valuable Plus for Business Visualization

    In some results, the user only asks for a web page without detailing the interactions, yet the model still adds features on its own, such as drag, rotate, annotate, focus, download, and image switching. This shows that current models are gradually acquiring a general structural understanding of “3D display applications.”

    For business deployment, this ability is valuable. In scenarios like property sand tables, city modeling, campus tours, and smart-community visualization, users often need not a static model but an interface they can explain, explore, and embed into business workflows.

    The typical structure of an interactive web page can be abstracted as the following pseudocode:

    1. scene = load_scene(model_file)
    2. render(scene, view="overview")
    3. on pointer_down:
    4. pause_auto_rotation()
    5. on pointer_move:
    6. rotate_scene()
    7. on pointer_up:
    8. resume_auto_rotation_after_delay()
    9. focus(target):
    10. camera.animate_to(target.bounding_box)
    11. highlight(target)
    12. show_label(target.name)
    13. show_annotation(node):
    14. display_panel(node.description)
    15. if user_idle(seconds=5):
    16. auto_rotate(scene)

    This pseudocode depends on no specific technology stack. It describes the general logic of 3D interactive display: load the scene, respond to drag, auto-rotate, focus on objects, show annotations, and control the viewpoint. What is really valuable is not any single feature but whether the model can combine these features into a complete, operable display flow.

    From a business perspective, this ability can cut a great deal of front-end development and 3D interaction configuration work. Scene nodes, viewpoint switching, annotation panels, and resource-loading logic that developers used to build by hand can now be generated as an initial version from a natural-language description, then fine-tuned and optimized by a person.

    Evaluation Dimensions for 3D Miniature Landscapes: More Than “Does It Look Right”

    Judging whether a 3D generation task passes should not rely on subjective impression alone. A more reasonable approach is to build a multi-layer evaluation system.

    1. Semantic Completeness

    Semantic completeness means whether the result covers the core theme and key elements in the prompt. For a detective agency theme, does it include the agency, cafe area, staircase, entrance character, and sign? If key elements are missing, even great visuals only count as a partial pass.

    2. Spatial Reasonableness

    Spatial reasonableness concerns whether the internal relationships are correct, including:

    • Whether the floors are clear;
    • Whether the staircase connections are sensible;
    • Whether the building proportions are balanced;
    • Whether character placement fits the scene;
    • Whether viewpoint switching feels natural;
    • Whether there are obvious intersections or floating objects.

    3. Interaction Usability

    Interaction usability mainly evaluates whether the web deliverable truly works, including:

    • Whether the model loads normally;
    • Whether dragging is smooth;
    • Whether auto-rotation is reasonable;
    • Whether annotations are accurate;
    • Whether focus lands on the target;
    • Whether download and display are complete.

    4. Render Quality

    Render quality affects the final reach and reporting effect, usually judged on:

    • Whether the composition is clear;
    • Whether the lighting is natural;
    • Whether materials are consistent;
    • Whether colors carry a sense of theme;
    • Whether the image is suitable for reports or presentations.

    5. Engineering Editability

    If the deliverable needs to enter a downstream production flow, its editability must be assessed:

    • Whether scene nodes are clear;
    • Whether objects can be selected individually;
    • Whether materials can be swapped;
    • Whether coordinates are standardized;
    • Whether it is convenient for secondary development;
    • Whether it can be imported into other 3D workflows.

    Seeing Model Evolution Through One Generation Task

    From a task like this, you can see that model capability is moving from “single-point generation” toward “flow delivery.” In the past, a model might produce only a piece of code, an image, or a text. Now it starts to attempt a complete delivery package that includes assets, renders, web pages, interactions, and a download entry.

    This shift means evaluation must be upgraded too. Instead of asking only “does it look right,” we should ask further:

    • Does it understand the task theme?
    • Does it plan a reasonable structure?
    • Does it fill in the key details?
    • Does it produce deliverable files?
    • Does it provide interaction?
    • Is it convenient for people to edit?
    • Can it enter a business system?

    For example, a missing sign is a small problem, but it exposes the model’s instability on key semantic elements. Conversely, the fact that annotation, focus, and drag guidance appear automatically in the interactive web page shows that the model has a strong structural understanding of 3D display scenarios. Whether a model is good often shows up in the difference between details it “fills in without being told” and details it “misses even when told.”

    Practical Advice for Business Deployment

    If you want to use this kind of 3D generation ability for real project delivery, you can approach it with engineering design in several ways.

    1. Build a Prompt Template Library

    Do not write prompts from scratch every time. Accumulate templates for different scenarios, such as:

    • Detective agency themes;
    • Commercial street miniature landscapes;
    • Smart campus visualization;
    • Property sand-table displays;
    • City landmark generation;
    • Exhibition space setup.

    Fixing deliverables, key elements, style, interaction requirements, and acceptance criteria in each template can noticeably improve generation stability.

    2. Introduce Staged Acceptance

    A 3D generation task is not suited to one-shot acceptance. It is better to split it into stages:

    1. Confirm the text understanding;
    2. Confirm a spatial structure sketch;
    3. Confirm the initial 3D asset;
    4. Confirm the render;
    5. Confirm the interactive web page;
    6. Confirm the final engineering delivery.

    This catches problems early and avoids the model continuing to fill in gaps in the wrong direction.

    3. Bring Generated Resources into Unified Asset Management

    If a model produces many resources, general cloud capabilities can help manage them. For example:

    • Use object storage to keep 3D models, renders, and web resources;
    • Use a content delivery network to improve cross-region access speed;
    • Use a logging service to record generation and access;
    • Use monitoring and alerting to catch resource-loading failures or API errors;
    • Use a container platform or serverless functions to host temporary preview services.

    These capabilities do not depend on any specific product; they are common building blocks of visualization-project engineering.

    4. Include Human Review in the Flow

    Generative models are good at producing a first draft but should not fully replace human judgment. Human review is especially important in cases like:

    • Key brand or business markers;
    • Projects where spatial proportions must match real scenarios closely;
    • Publicly released display pages;
    • Content involving compliance, copyright, or data accuracy;
    • Systems that must embed into a client’s business workflow.

    Summary: The Core of 3D Generation Evaluation Is “Complete Delivery”

    Overall, current LLMs in 3D miniature landscape generation are no longer limited to simple modeling. They are gradually covering semantic understanding, spatial reasoning, asset generation, render output, and interactive web pages. When evaluating this ability, you should not just check whether the model produced a scene that looks polished; you should focus on whether it completed a full delivery.

    A high-quality generation result should at least meet these conditions:

    • Clear theme;
    • Reasonable structure;
    • Complete key elements;
    • Consistent style;
    • Clear renders;
    • Usable interaction;
    • Editable project;
    • Easy to optimize further.

    Take the detective agency miniature landscape as an example: even if the staircase, floors, characters, and cafe area are generated, a missing sign still means it is not a complete delivery. What is missing is not an ordinary decoration but the core semantic element of theme recognition.

    Conversely, if the model can automatically add web interactions, scene annotations, viewpoint focus, and drag guidance under a limited prompt, it shows that it is starting to develop delivery awareness for business scenarios. In the future, the value of this ability lies not in whether a single generation is perfect, but in whether prompts, acceptance criteria, and human collaboration can reliably turn generation results into editable, presentable, deliverable digital assets.