LLM 3D Creation Test: From Prompt Design to Deliverable Evaluation

    This article reports a hands-on evaluation of an LLM’s 3D creation ability using a bare prompt to generate a voxel diorama of the Mouri Detective Agency in Blender. It examines structural versus semantic fidelity, the model’s unexpected self-extension of an interactive web page, a reusable four-step evaluation method, and the quota and cost constraints that affect practical testing.

    1. Background: Why 3D Creation Became a New Evaluation Direction

    Over the past year or so, community evaluation of large model capabilities has shifted noticeably. Early tests focused mainly on code generation and logical reasoning, then moved toward front-end page reproduction. More recently, 3D modeling and spatial reasoning have become a new bellwether.

    Several technical drivers are worth analyzing:

    1. Code capability is becoming more homogeneous. The gap between mainstream models on routine coding tasks is shrinking. Simply comparing whether a model can write runnable code is no longer enough to distinguish capability levels.
    2. 3D tasks are stronger comprehensive tests. A single 3D miniature landscape involves spatial geometry understanding, proportion control, materials and color schemes, and semantic scene reproduction. Weakness in any one of these areas shows up in the final deliverable.
    3. The delivery format is closer to real production. 3D creation usually requires an editable project file such as a Blender project, a rendered image, and an interactive web preview. This maps directly to the dual delivery requirement in real projects: source file plus presentation layer.

    Some model vendors also promote 3D creation as a headline capability and provide dedicated product entry points that support exporting the same model as a modeling project, a web page, or a printable file. This makes horizontal hands-on testing more accessible for ordinary developers.

    2. Test Design: Why Use a Bare Prompt

    This test deliberately did not attach any external skills, retrieval augmentation, or workflow orchestration. It used a single original prompt connected directly to the model. There are three reasons:

    • Eliminate interfering variables. When skills are attached, it is difficult to tell whether the final result comes from the model itself or from external tools.
    • Expose real weaknesses. Under a bare prompt, the model must independently handle understanding, planning, and execution, so defects appear more directly.
    • Approximate the minimum usage cost. Most users start with a single bare prompt when they first try a model. This represents the lower bound of capability.

    The prompt used was:

    1. Create a 3D voxel diorama themed around the "Mouri Detective Agency" in Blender.
    2. Inspired by the detective office from the classic anime, make it the visual highlight.
    3. Express it in a voxel-block language, and deliver an editable Blender project and a render.

    Note that this prompt specifies only three things: the theme, the visual style (voxel), and the deliverables (project file plus render). It deliberately does not mention an interactive web page, an annotation system, or other additional features. This point is valuable in the later evaluation.

    3. Deliverable Evaluation: Two Sides of the Model’s Reproduction Ability

    3.1 The Model Itself: High Structural Fidelity, but Semantic Omissions

    The final output included two model render PNGs and one Blender project file. In terms of structural completeness, the result exceeded expectations. The building’s floor structure, the staircase to the second floor, and the first-floor spatial zones were basically built. Two character models were even placed at the entrance as accents.

    The problems were also typical. The highly recognizable signboard on the wall in the original anime was missing from the model. This is the visual anchor of the entire piece, and its absence greatly reduces overall recognizability.

    This exposes a general pattern: current models reproduce “structural elements” in a scene relatively well, such as stairs, floors, doors, and windows—objects with clear geometric rules. But they tend to miss “semantic elements,” such as signboards and posters that carry narrative information. The former can be inferred from common sense; the latter requires the model to genuinely “understand” the cultural context of the theme.

    Another detail is deformation in character modeling. The hairstyle of one character was clearly distorted and almost unrecognizable. This shows that in low-poly or voxel styles, the model’s ability to handle organic forms—people and curved silhouettes—is still noticeably weaker than its handling of regular geometric objects.

    3.2 Interactive Web Page: Unexpected “Self-Extension” Capability

    More interesting was the second follow-up prompt, which only asked to open the model in a web page and did not mention any functional requirements. The resulting web page included a whole set of features beyond expectations:

    • Basic interaction: model dragging and auto-rotation
    • Scene annotation system: add explanatory labels to elements in the model
    • View focus: a sidebar lists three key elements in the scene, and clicking one flies to the corresponding view
    • Deliverable integration: a download entry for the model and a toggle between two renders
    • Onboarding guidance: drag-operation guidance shown on first open

    This phenomenon—features not requested in the prompt but proactively completed by the model—can be called self-extension of the deliverable. It reflects that during training, the model formed a strong prior about what a complete 3D presentation web page should look like.

    3.3 Engineering Value of Self-Extension

    For developers, this capability has direct value in specific business scenarios. Take smart visualization projects—real estate sandboxes, park modeling, and smart city dashboards. Scene annotation and view focus are exactly the kind of must-have features these projects require:

    1. Typical requirement mapping:
    2. - 3D sandbox model display -> drag / auto-rotate / view focus
    3. - Building information descriptions -> scene annotation system
    4. - Demo and reporting scenarios -> preset views + auto cruise
    5. - Internal review delivery -> source file download + render archive

    If these features had to be written manually, a front-end solution based on something like Three.js would require at least hundreds of lines of code. The model fills them in at once, greatly shortening the distance from “receiving the model” to “having something demonstrable.”

    4. A Reusable Evaluation Methodology

    Based on this hands-on test, I summarized a four-step method for evaluating an LLM’s 3D creation capability:

    Step 1: Use a bare prompt to test the lower bound of capability. Design a theme with a clear “standard answer,” such as a well-known scene or landmark building, so that reproduction fidelity can be judged. The prompt should specify only three elements: theme, style, and deliverables. Leave everything else to the model.

    Step 2: Check reproduction fidelity at both the structural and semantic levels. At the structural level, check geometric completeness such as proportions, floors, and passages. At the semantic level, check whether iconic theme elements are preserved. Semantic omissions are a common weakness of current models and the main source of point deductions.

    Step 3: Observe self-extension in the deliverable. Compare the functions required by the prompt with the functions actually delivered. If the difference is positive—the model did more—and the extra functions are high-value, the model has solid prior knowledge. If the difference is negative—the model did less—there is an instruction-following problem.

    Step 4: Verify engineering usability. Open the Blender project and check whether the model is editable, whether the hierarchy is reasonably named, and whether materials are independently adjustable. A “dead model” that can only be viewed but not modified has no value in real production.

    5. Quota and Cost: Practical Constraints in Evaluation

    There is another easily overlooked issue in evaluation: usage quota. In the early release period of a new-generation model, the community commonly reports that free quota is consumed quickly and that complex 3D tasks use a large number of tokens per generation, making it easy to hit the limit. Models accessed through third-party platforms also have daily limits. This means:

    • When conducting systematic evaluations, quota planning should be included in the test plan, with priority given to core scenarios.
    • For enterprise users, the per-generation cost of 3D generation tasks is significantly higher than that of text tasks, so cost estimation must be done before launch.
    • It is advisable to record the consumption of each call in evaluation reports to facilitate horizontal comparison of cost-effectiveness.

    6. Conclusion

    The conclusion of this hands-on test can be summarized in one sentence: 3D structural reproduction has reached a usable level, and the self-extension capability of the interactive web page is surprising, but semantic element reproduction and character details remain clear weaknesses.

    For developers, a practical approach at the current stage is to treat the 3D scene generated by the LLM as a “high-completion draft,” then manually fill in iconic details such as signboards, textures, and character tweaks, and finally use its automatically generated interactive web page directly for demo and reporting. This collaborative model—“the model builds the skeleton, the human adds the soul”—may be the right way for 3D generation capability to enter the real production chain. I will later test more complex real estate sandboxes and smart city scenes and share further observations then.