LLM-Generated 3D Voxel Street Scenes: A Guide to Prompt Design and Acceptance
This article looks at how a large language model (LLM) handles 3D content generation through a single case study: prompting an LLM to build a detective-agency-themed voxel miniature street scene in Blender. It explains why bare prompts, with no templates or examples, expose a model's true spatial-reasoning ability, then evaluates the output across theme consistency, element completeness, spatial structure, character modeling, render quality, and interaction. The scene reconstructed the overall structure well but dropped its iconic sign, showing why "looks right" is not the same as "all elements present." Readers get a six-part prompt template, a reusable acceptance checklist, and a generate–accept–fix workflow, plus a look at how the model's self-added interactive webpage (scene annotation and view focusing) can transfer to real estate sandboxes and city modeling. The core takeaway: prompts set the lower bound, while acceptance criteria set the upper bound.
1. Introduction: The Evaluation Focus Shifts from “Can It Write Code?” to “Can It Build a Scene?”
For a long time, the standard for measuring an LLM’s engineering ability was whether it could produce runnable code. Over the past year, 3D content generation has gradually entered mainstream evaluation, and related tasks now ask models to combine modeling scripts, web interactions, and other means to turn a piece of natural-language description into a renderable, editable scene.
The reason is easy to understand: a 3D scene simultaneously tests spatial understanding, element organization, and detail modeling. The model must read spatial relationships like “a second-floor office, a first-floor café,” translate them into coordinates, dimensions, and orientation, and make tradeoffs between abstraction and realism, keeping the subject recognizable without dropping any key element.
This article follows one concrete case: using a bare prompt, having an LLM create a detective-agency-themed voxel miniature street scene. The scene is drawn from a classic setting in a detective anime: the second floor facing the street is the detective agency, the first-floor shopfront is a café, and an external staircase on the side leads up to the second floor. The main characters are Kogoro Mouri, who lives at the agency, and the high-school detective Shinichi Kudo, who has been turned into a child and is staying at his home. The original also mentions that in the anime, a prominent “Mouri Detective Agency” sign hangs on the agency wall.
The final deliverables in the case included an editable project file, two rendered images, and an interactive webpage added after a manual follow-up request. The overall structural reconstruction was decent, but the sign was entirely missing. The final conclusion was “the reconstruction is okay, but it can’t get full marks.” This article is not a product review; it treats the case as a sample and distills three main threads: how to design the prompt, how to locate and fix defects, and how to accept the generation results. All views come from a single experiment and are therefore case-specific.
2. Experiment Design: Controlling Variables with a Bare Prompt
A “bare prompt” refers to a natural-language instruction given directly to the model without any external assistance. The “external assistance” here includes templates, few-shot examples, project specifications, domain knowledge, and callable skill plugins. This design was chosen for three reasons:
- Removing engineering safety nets: With scaffolding or templates, a model can always produce something that “looks the part.” Only a bare prompt exposes the model’s true level of spatial understanding and requirement inference.
- Closer to real needs: Many requirements come from situations where “there’s a picture in someone’s head,” and being able to describe the theme and rough structure clearly is already difficult.
- Easy to reproduce: The input is a single piece of text, so any reader can rerun the same experiment on a comparable platform.
The full prompt used in the experiment was:
Please use Blender to create a 3D voxel miniature landscape themed on the "Mouri Detective Agency."Inspired by the Mouri Detective Agency from Detective Conan, make the Mouri Detective Agencythe visual highlight, render it in a voxel-block style, and deliver an editable Blenderproject and rendered images.
A quick breakdown shows this prompt actually conveys only five pieces of information:
- Tool: specifies the modeling software;
- Theme: a detective agency, with the source of the work given;
- Style: voxel-block language;
- Focus: the agency is the visual highlight;
- Deliverables: an editable project and rendered images.
It does not specify the number of floors and the street-facing relationship, the staircase location, the type of ground-floor shop, whether characters are needed, or what viewing angle to use, nor does it mention any web requirement. Interestingly, a final follow-up added “open the file in HTML,” yet the model delivered an interactive webpage. From this we can see that the model has its own understanding of the “HTML” medium. This shows that the more vaguely the delivery format is expressed in the prompt, the more room the model has to improvise, and the less predictable the result becomes.
It is also worth noting that some platforms set a daily quota on model calls, so continuous, repeated experiments will be limited by this. Therefore, the prompt should be written clearly in one pass, with fewer trial-and-error detours.
3. Accepting the Generation Results: Complete Structure, Missing Key Elements
Looking at the results, this generation was “similar in form but incomplete in spirit.” It should be emphasized that the voxel style is itself a highly abstract expressive language and should not be held to demanding standards of realistic detail. But for a task “themed around a detective agency,” the iconic sign is precisely the core element used to “make the theme explicit.”
1. Spatial structure: complete skeleton
- An external staircase on the side leading to the second floor;
- The detective agency on the second floor;
- The café as the first-floor shopfront;
- Two characters standing at the door (a young boy and an adult woman).
From this brief prompt alone, the model “remembered” the typical composition of the scene from its training data and did not draw the key architectural relationships incorrectly. This shows that for well-known subjects, the model can draw on existing knowledge to fill in the implicit scene structure, provided these structures have sufficient density of corresponding descriptions in the corpus.
2. Iconic sign: the most prominent element is absent
The very prominent “agency” sign from the anime was not modeled, for three possible reasons:
- High text-geometry cost: If the sign contains Chinese or Japanese characters, the glyphs must be converted into 3D geometry. After voxelization, the strokes blur together at a limited resolution.
- Insufficient semantic association: The concept of “agency” the model learned may lean more toward structural features like doors, windows, desks and chairs, and floors; the “sign” was not strongly associated with the theme as a core visual feature.
- Treated as decoration rather than a theme element: In the model’s generation strategy, the sign belongs to optional decoration.
For the absence of a “landmark/sign” element, three approaches can verify it:
- Repeated sampling: Generate multiple times with the same prompt. If most results lack the sign, this indicates a stability problem rather than just random sampling bias.
- Fixed random seed: Repeatedly generate in a modeling-script environment that supports a fixed random seed, and observe whether the sign is missing.
- Adding constraints in follow-up: Explicitly require “a sign reading ×× must be placed on the street-facing wall.” If the sign can be added at this point, it means the model is capable of generating it, but the original prompt did not treat it as a required element.
Based on the principle that “the core identifier of the theme must appear,” the result cannot get full marks. But the fix is low-cost and can be resolved in one iteration of multi-round generation; the specific approach is given later.
3. Character modeling: recognizability relies on context rather than the model itself
The adult female character’s hairstyle is fairly abstract and hard to identify on its own. This is less an error than an inherent tradeoff of the voxel style: low-resolution blocks cannot reproduce the outline of long hair or a ponytail, so recognition can only rely on scene context (a female character standing at the agency door) and additional labeling. If the task requires characters to be recognizable, constraints such as head proportions and hairstyle outline should be given separately in the prompt, or this tradeoff should be explicitly accepted during acceptance.
4. Rendered images: acceptable presentation
The two rendered images have complete composition, unified color tones, and a good filter look, with no problem as a presentation of results. Two points need attention:
- Whether the default viewing angle of the rendered images best showcases the theme should be judged by core elements such as the sign, staircase, and storefront;
- Promotional materials should note that the images have been filter- or color-processed, to avoid a gap with the appearance of the model itself.
4. Interactive Webpage: Filling In Functionality Beyond the Prompt
The most “unexpected” gain this time was the webpage. The prompt said nothing about a webpage, and only a later line added “open the file in HTML.” Yet the interactive webpage the model delivered proactively added many features the designer never spelled out.
1. Basic interactions
- Mouse dragging: You can freely rotate and zoom to observe the model;
- Auto-rotation: Auto-rotation is on by default, making it easy to browse the whole thing;
- First-drag guidance: On first open, a drag hint is given, solving the “didn’t know it could rotate” problem, a common and thoughtful practice in interaction design.
2. Scene annotation
The webpage can place clickable annotation points in the scene, providing text descriptions for the key elements in the model (staircase, agency, café). This capability has clear engineering value:
- In visualization projects such as real estate sandboxes and city modeling, annotation is key to “explaining” a 3D model clearly;
- Annotation points usually carry bindable text or data fields, suitable for business information such as area, floors, and shop type;
- It upgrades “the model looks good” to “the model is readable and explainable.”
3. View focusing
A sidebar lists the three focusable elements in the model; clicking one smoothly moves the camera to the corresponding view.
- It solves exactly the pain point of voxel scenes: once there are many elements, manual dragging makes it hard to align on target in one go.
- Annotation answers “what is this,” focusing answers “where is it and what does it look like”; the two together noticeably improve the completeness of the interaction experience.
4. Centralized display of outputs
The webpage integrates features like downloading the model and switching between the two rendered images, bringing the project file, rendered images, and interactive preview together under one entry. This is very delivery-friendly: the user does not need to open the modeling software, an image viewer, and a webpage separately.
These features share one thing in common: none of them were mentioned in the prompt. Most of them “translated” common interaction habits from modeling-software interfaces into the webpage. This means the model brought the interaction paradigm of general-purpose visualization applications into the generated webpage. This reflects its requirement-inference ability, but the inferred features may not fully match the user’s intent, so the interactive webpage also needs to be accepted separately.
5. Prompt Design: Putting Key Elements First
The fundamental problem this experiment exposed is that the prompt “wrote the theme specifically, but the elements not specifically.” “Themed around a detective agency” only clarified the subject but did not clarify what must appear, where it goes, and in what form it appears. To address this, you can use a “six-part” prompt template:
[Theme and inspiration] State in one sentence what the scene is drawn from and what you want the viewer to see at a glance.[Spatial structure] Number of floors, functional zones (e.g., ground-floor shopfront / second-floor office), and the main entrance and staircase locations.[Required-elements list] List the non-omittable elements one by one, especially the theme landmark and its position and size.[Style and granularity] Voxel-block size (e.g., 0.2m per grid), overall color tendency, and details to emphasize.[Deliverables] Editable project file, rendered images with a specified count and viewing angles; if a webpage is needed, write out the interaction requirements.[Prohibitions] Specify elements, styles, or effects that are not wanted, to avoid deviations from the model's free improvisation.
This template has three key points:
- Structure first, details second. The model builds the skeleton first, then fills in details, avoiding a locally refined but globally unbalanced result.
- List landmarks separately. Elements like signs, doorplates, and theme walls form their own line, with “must appear” wording.
- Write out all deliverables. List the project file, rendered images, and interactive webpage separately. If a webpage is needed, state directly functions such as annotation, focusing, and auto-rotation.
The template solves the “get it right the first time” problem, but the randomness of single-shot generation still exists and needs to be used together with an acceptance checklist.
6. Advanced Use of the Interactive Webpage: A Real Estate Sandbox Driven by Scene Annotation
The interactive webpage is not just a one-off presentation of results. Extracting the two capabilities, scene annotation and view focusing, forms a minimal implementation of a “real estate sandbox”:
- Annotation layer: annotation points as business-info anchors for buildings and unit types, associated with fields like building number, unit area, and orientation;
- Focus layer: camera views as reusable presets, supporting progressive drill-down from “overview, zone, single building”;
- Interaction layer: dragging, zooming, and auto-rotation form the basic browsing experience, with first-open guidance lowering the barrier to use.
The reasoning is as follows:
- The basic functions of the interactive webpage (camera control, hotspot annotation) are highly isomorphic with sandbox business needs;
- You only need to replace decorative annotations like “staircase, agency, café” with business annotations like “building, unit type, orientation”;
- A one-off miniature street scene can thus evolve into a sustainable visualization tool.
For a team, what matters more is accumulating a reusable scene-annotation configuration structure, for example:
// Illustrative: annotation point configuration (fields can be replaced per business scenario)const annotations = [{id: "tower-a",label: "Building 1",position: [0, 6, 0], // Position of the annotation point in world coordinatesdata: { floors: 18, units: 216 }},{id: "amenity-center",label: "Community Center",position: [-8, 1.5, 4],data: { area: 1200 }}];// Click annotation → camera smoothly focuses on the target position and displays business infofunction focusOn(target) {camera.lookAt(target.position);showPanel(target.data);}
Such annotation configurations are themselves version-manageable assets: the model generates the scene, and humans calibrate the annotations and business fields, each doing their part.
7. A Reusable Acceptance Checklist
To avoid “scoring by feel,” you can build an acceptance table by dimension, checking each item and drawing a conclusion:
| Dimension | Checkpoint | Conclusion in This Case |
|---|---|---|
| Theme consistency | Does the theme landmark appear, and is its position reasonable | Sign missing; fail |
| Element completeness | Are all required elements in the prompt present | Staircase, agency, café, and characters all present |
| Spatial structure reasonableness | Do the floor relationships and circulation follow real-world logic | External staircase, ground-floor shopfront, and second-floor office match the original |
| Modeling reasonableness | Are there obvious distortions in character proportions, hairstyle, etc. | Female character’s hairstyle is abstract; recognizability is insufficient |
| Render quality | Can composition, tone, and viewing angle highlight the theme | Good overall look; viewing angle could be improved |
| Interaction functionality | Are dragging, rotation, annotation, focusing, and guidance usable | Exceeded expectations; functionality complete |
| Project editability | Can it be opened in modeling software and further modified | Editable project file delivered |
| Requirement-inference reasonableness | Do the self-added features fit actual use scenarios | Matches visualization project needs |
Two notes:
- Weights can be adjusted per task. For promotional materials, render quality carries more weight; for visualization systems, interaction and annotation carry more weight.
- The total score is jointly determined by two core metrics: first, whether the theme landmark appears (keyword or visual similarity), and second, whether key interactions are usable. If either is missing, even if the whole “looks good,” it should not get a high score.
8. From Single-Point Generation to an Engineering Workflow
Stringing together the above steps forms a closed loop of “generate, accept, fix”:
- Write the prompt clearly: Use the six-part template to write out the structure, elements, style, and deliverables in one pass.
- Generate and accept item by item: Check against the acceptance checklist, focusing on the landmark and interaction.
- Add one round targeting the defect: For example, add a prompt like “add a sign next to the second-floor window on the street-facing wall, presented in block characters or simple geometry, with a prominent size”; or manually add it in the modeling software.
- Accept the interactive webpage separately: Check whether the annotation text, focus views, and guidance copy match the scene.
- Accumulate assets: Organize the prompt template, acceptance checklist, and annotation configuration structure into reusable team assets, reducing the quota consumption from repeated experiments.
Two boundaries need to be stated objectively:
- Stability of generation results: Even with a very detailed prompt, randomness can still cause detail deviations, and manual calibration cannot be entirely omitted.
- Copyright and compliance: When creating derivative works from existing works such as animation and film for public release or commercial use, you need to pay attention to the original work’s copyright and usage licensing. When real people’s portraits or building exteriors are involved, corresponding assessment should also be done.
9. Conclusion
Back to the opening question: how should we view LLM performance on 3D miniature landscapes? This experiment gives a two-sided answer.
On one hand, the overall spatial structure was reconstructed quite well. The model can reconstruct key components like the staircase, agency, and ground-floor shop from a single sentence, showing that its structured understanding of classic scenes already has practical value.
On the other hand, the theme landmark was missed. This precisely shows that “looking like it” does not equal “all elements present,” and that acceptance cannot rely only on the first impression.
The most valuable gain was actually the interactive webpage that was not in the prompt but that the model completed proactively. The two capabilities, scene annotation and view focusing, can be directly transferred to visualization scenarios like real estate sandboxes and city modeling. They remind us: the prompt only determines the floor; the acceptance criteria determine the ceiling. Next time you have a model create a 3D street scene, try first writing “where the sign goes and what it looks like” into the prompt, then go through the acceptance checklist item by item.
