Evaluating AI 3D Scene Generation: A Bare-Prompt Testing Approach and Best Practices
This article examines how to evaluate large language models' ability to generate 3D scenes. Using a bare-prompt test that asked a model to build a voxel-style detective agency miniature scene—without reference images, extra skills, or component libraries—it shows why acceptance criteria should go beyond visual similarity to cover semantic completeness, spatial rationality, engineering editability, rendering quality, and interactive completeness. It also breaks down the interactive web output, proposes a five-part prompt structure (Goal–Objects–Structure–Interaction–Acceptance), presents a multi-dimensional scoring system, and outlines how to turn one-off demos into a reusable evaluation pipeline.
1. From Code Generation to 3D Creation: The Evaluation Focus Is Shifting
In the past, code generation was often the most intuitive way to observe a large model’s capabilities. But as models grow stronger at understanding multimodal structure, spatial relationships, and engineering-ready delivery, 3D scene generation is becoming an important benchmark for overall capability. Compared with a single block of code, 3D scene generation involves semantic understanding, spatial layout, object organization, material expression, interaction design, and file delivery all at once. It is much closer to the full “from requirements to a usable prototype” workflow found in real business scenarios.
In one bare-prompt test, the requirement was to generate a 3D miniature scene on the theme of a “detective agency.” The prompt came with no reference image, loaded no additional skills, and specified no complex component libraries or plugins. The model was simply asked to generate a voxel-style scene based on the theme and to deliver an editable 3D project, rendered images, and an interactive web page. The purpose of this test was not to verify whether the model can use a specific tool, but to observe its comprehensive generation ability under open conditions: Can it break a single sentence down into scene objects? Can it organize those objects into a reasonable spatial structure? Can it fill in implicit delivery requirements and produce output that can be further refined?
2. The Value of Bare-Prompt Testing: Reduce External Aids, Observe Real Capability
When evaluating a model’s 3D generation ability, a common mistake is to package the task too completely—for example, by presetting a detailed workflow, providing a reference image, specifying a component structure, or even requiring the model to follow a fixed template. These approaches are closer to “toolchain collaboration.” They can improve the final result, but they make it hard to reflect the model’s own spatial-understanding ability under open conditions.
The core of bare-prompt testing is to observe the model’s default behavior with minimal context. At a minimum, it can test three levels of capability:
- Semantic extraction: Can the model identify the necessary objects from the theme—such as buildings, signs, stairs, entrances, characters, and furniture?
- Spatial organization: Can the model judge the relative positions of objects, floor relationships, circulation paths, and visual focal points?
- Delivery completion: Based on requirements like “editable project, rendered images, interactive page,” can the model proactively organize a reasonable output structure?
This kind of test is not suitable as the sole scoring standard, because it may expose a model’s weaknesses under complex constraints. But it is very well suited for observing a model’s baseline capability, default tendencies, and weak spots. In particular, when a model fills in some reasonable functionality without being explicitly asked to, it often reflects its underlying task understanding.
3. 3D Generation Acceptance Is Not Just About “Does It Look Right”
From the delivered results of this test, the model had completed most of the basic spatial structure, including a side staircase, a second-floor agency, a ground-floor reception area, and a character at the entrance. This shows that the model has a certain scene-organization ability: it can extract the relationship of “building—floor—functional area—accessories” from a text description and convert it into a 3D scene.
However, 3D generation should not be judged by visual similarity alone. A more reasonable acceptance approach should cover the following five levels.
1. Semantic Completeness
Semantic completeness focuses on whether all the core elements explicitly required in the requirement are present. In this case, the model missed the most critical element: the “agency sign.” A sign is not just a decorative element—it is a core semantic anchor for theme recognition. If this element is missing, then even with a fairly complete building structure, the result should be considered only partially acceptable.
Therefore, in scoring, a “core identifier” should be treated as a high-weight item rather than an ordinary accessory. For commercial spaces, stores, exhibition halls, campus entrances, and similar scenes, identifiers often carry the theme-recognition function and must be listed as separate acceptance points.
2. Spatial Rationality
Spatial rationality focuses on whether the position, proportion, occlusion, and circulation relationships of objects conform to reality or the specified style. For example: Does the staircase connect to the correct floor? Is the character at the entrance in a reasonable position? Do the doors and windows match the building structure? Can the second-floor space be reached naturally from the entrance or the stairs?
In voxel or block-style scenes, proportions may be somewhat exaggerated, but basic spatial logic should still hold. Otherwise, the scene looks like a scattered pile of objects rather than a coherent miniature space.
3. Engineering Editability
The value of 3D generation is not just producing an image, but producing a project that can continue to be edited. Editability includes whether object names are clear, whether grouping is reasonable, whether materials are adjustable, and whether models are easy to replace or extend later.
If the internal structure of the generated project file is chaotic—with all objects mixed together and no way to quickly modify a single object—then even with good visual results, it will affect later use. Especially for scenes delivered to design, product, operations, or visualization teams, engineering editability is often more important than a single render.
4. Rendering Expression
A rendered image is a comprehensive reflection of the model’s understanding, materials, lighting, and composition. When evaluating rendered images, you can look at whether the colors are consistent, whether the shadows are natural, whether the subject stands out, and whether the viewing angle shows the scene’s features. For a miniature scene, top-down, oblique, and close-up views all help judge how well the model grasps both the overall scene and its details.
5. Interactive Completeness
The interactive web page was one of the more impressive parts of this test. The prompt only asked to open the file in a web page, but the model additionally generated features such as drag, auto-rotation, scene annotation, view focusing, model download, image switching, and an operation guide. Although these capabilities were not explicitly requested, they happen to form an important part of the user experience of a 3D deliverable.
For smart-visualization projects, interactivity often determines whether a 3D model can be understood by non-technical people. A web page that can rotate, zoom, click annotations, and focus on objects is easier to use for reporting, presentation, and training than a static render.
4. Interactive Web Capability: From Model Delivery to a 3D Application Prototype
In this test, the interactive web page generated by the model included several features worth noting:
- The 3D view supports drag and auto-rotation;
- It provides scene-object annotations;
- The sidebar supports focusing the view by object name;
- It offers a model download entry;
- It supports switching between rendered images;
- It shows an operation guide on first entry.
These features show that the model was not just generating a 3D asset, but was also trying to organize a small 3D application. If abstracted as an engineering structure, it can be broken down into the following modules:
- Scene layer: Responsible for model loading, camera control, lighting, and rendering.
- Annotation layer: Maintains object names, descriptions, coordinates, and display rules.
- Interaction layer: Handles rotation, zoom, focusing, layer switching, and download.
- Presentation layer: Provides introductory copy, image previews, and operation tips.
This structure has practical value for later extension. For example, in scenarios such as real-estate sandbox models, campus displays, city modeling, equipment descriptions, and exhibition-hall tours, scene annotation can turn “visual objects” into “searchable, locatable, explainable information nodes.” Users are no longer just looking at a 3D model—they can click a region and view its name, function, or status.
Especially noteworthy is the “view focusing” capability. If the sidebar lists the key elements in the model and lets users click to automatically locate the corresponding object, it can significantly lower the cost of understanding a complex scene. For multi-story buildings, complex equipment, or multi-zone campuses, this capability is very practical.
5. Improving Bare Prompts: Make Implicit Acceptance Criteria Explicit
Although the original prompt gave the theme, style, and deliverables, it was still not clear enough about the priority, quantity, and naming rules of key elements. For the model, “agency” may have been understood only as a building block, while strongly associated semantics such as signs, door plates, and markers need to be listed explicitly.
To improve the stability of 3D generation results, it is recommended to use a five-part prompt: “Goal–Objects–Structure–Interaction–Acceptance.”
Example Structure
Goal: Generate a voxel miniature scene on the theme of a detective agency.
Objects: Must include the agency’s main body, a street-facing sign, a staircase, a second-floor office area, a ground-floor reception area, a character at the entrance, tables and chairs, a filing cabinet, and doors and windows.
Structure: The building has two floors; the staircase is on the side; the agency entrance faces the street; the sign is placed in a prominent position on the second-floor exterior wall.
Interaction: Deliver an interactive page supporting rotation, zoom, scene annotation, object focusing, and download.
Acceptance: All core objects must appear and be clearly named; the sign’s text or graphics must be recognizable; rendered images must show front and side views.
The essence of this approach is to move the acceptance criteria to the front. When generating 3D content, a large model is not simply “drawing”—it is performing constraint solving. It needs to judge, within limited context, which objects must be created, which can be omitted, and which need to be emphasized. The clearer the acceptance criteria, the fewer semantic omissions there will be.
6. Build a Multi-Dimensional Evaluation System
To make a horizontal comparison of 3D generation ability across different stages and scenarios, you can build a quantifiable or semi-quantifiable evaluation table.
1. Semantic Match
Check whether all specified objects are covered, whether irrelevant objects appear, and whether the theme and style are understood. For example, a detective agency scene should reflect office, reception, signage, and staircase elements—not just an empty building.
2. Spatial Structure Rationality
Observe whether the floor relationships, staircase connectivity, entrance orientation, character proportions, and occlusion relationships are natural. For multi-floor scenes, pay special attention to whether there is a reasonable circulation path between the upper and lower floors.
3. Key Element Recognition
Focus on whether strongly semantic elements such as signs, markers, entrances, and staircases are placed accurately. These elements often determine whether the scene’s theme holds up and should be core scoring items.
4. Editable Engineering Quality
Evaluate whether object names are clear, whether grouping is reasonable, whether materials are adjustable, and whether the model is easy to modify further. A high-quality 3D project should let later team members quickly locate objects, modify materials, replace assets, or extend functionality.
5. Rendering Output Quality
This includes composition, color, shadows, sharpness, and style consistency. Rendered images are not only for display but can also verify whether the model understands spatial relationships and material expression.
6. Interaction Experience
Check whether the page loads stably, whether rotation is smooth, whether annotations are accurate, whether focusing lands correctly, and whether download works. For an interactive web page, stability is often more important than the number of features.
7. Robustness
Under different prompt wordings, different theme complexity, and different interaction requirements, check whether the output stays stable. A model’s capability should not show up in only one successful case; it should maintain relatively consistent quality across multiple tasks.
8. Resource Consumption and Quota Management
For actual platform calls, record the time taken per task, the number of retries after failures, output file size, and quota consumption. If the daily quota is limited, prioritize small scenes for smoke testing, then gradually scale up to complex scenes.
7. From a Single Demo to a Reusable Workflow
A single generation result only reflects one incidental performance. To form a stable capability, you need to consolidate the workflow.
1. Task Grading
3D generation tasks can be divided into three levels:
- Simple tasks: A single object or a single room, such as a table or an office.
- Medium tasks: Multi-story buildings with several functional areas, such as the detective agency in this case.
- Complex tasks: Blocks, campuses, city sandbox models, or 3D visualizations with business data.
Different levels of tasks should use different prompt complexity and acceptance criteria. Simple tasks are suited for quickly verifying a model’s baseline ability; medium tasks are suited for verifying spatial-organization ability; complex tasks are suited for verifying system stability and engineering capability.
2. Layered Deliverable Verification
It is recommended to verify in stages:
The first stage only verifies the model’s geometry and key semantics. The second stage verifies rendered images and material style. The third stage verifies web-page interaction and annotation information.
This lets you catch problems early and avoid discovering a missing core object only after the interactive page has been generated. Especially for complex scenes, verifying the object list and spatial layout first, then optimizing the visuals and interaction, can significantly reduce cost.
3. Archiving Results
Save the prompts, output models, rendered images, interactive pages, and evaluation records together for review. This can be combined with common cloud capabilities such as object storage, logging services, and monitoring and alerting to form a traceable generation pipeline. For example, on Baidu AI Cloud infrastructure, you can store model files and web assets in object storage, write call records to a logging service, and use monitoring and alerting to observe task failure rates and resource consumption.
4. Automated Evaluation
For high-frequency scenarios, build a fixed test set that includes the theme name, required-object list, interaction items, prohibited items, and output format. Every time the model or prompt changes, run a regression test to avoid capability drift. For enterprise-internal scenarios, you can also incorporate business terminology, spatial naming, and presentation standards into the evaluation template so that the results are closer to real usage needs.
8. Common Problems and Correction Strategies
In actual testing, the common problems fall mainly into the following categories.
1. Missing Key Elements
The fix is to break objects down to the smallest granularity and specify their position and importance. For example, instead of just writing “agency,” write “a detective agency sign must be placed on the second-floor exterior wall.”
2. Unstable Style
The fix is to add constraints on materials, colors, lighting, and block granularity. Although voxel style is simple, it still needs to specify whether curves are allowed, whether a low-poly feel is required, and whether a consistent color scheme is needed.
3. Incorrect Spatial Relationships
The fix is to describe the floors, entrance, staircase, and viewing direction, rather than just giving a theme. For example, “the staircase is on the right side of the building, connecting the first and second floors, and the entrance faces the main street” helps the model understand far better than “generate a building.”
4. Too Much or Too Little Interaction
The fix is to specify which features the page must and must not include. If only a basic preview is needed, you can avoid generating overly complex interactions; if a business presentation is needed, you should explicitly require annotation, focusing, download, and guidance.
5. Unusable Project File
The fix is to require object naming, grouping, editable parameters, and a unified coordinate system. For later development, a clear project structure is often more critical than visual detail.
6. Single Rendered View
The fix is to require front, side, top-down, or key close-up views. Multiple rendered angles help judge whether the model truly understands the spatial structure.
9. Conclusion: 3D Generation Enters the “Comprehensive Delivery” Stage
This bare-prompt practice shows that large models have moved from “can they make a model” to “can they deliver a usable 3D experience.” The evaluation focus should shift from pure graphic generation to a comprehensive examination of semantic completeness, spatial rationality, engineering editability, and interactive completeness.
For developers, a bare prompt does not mean fewer constraints are always better—it means being clear about which constraints must be written into the prompt. For enterprise users, the value of 3D generation lies not only in automatically generating assets, but more in quickly forming a business prototype that can be displayed, annotated, downloaded, and extended. Only by combining evaluation criteria, prompt structure, and engineering acceptance can you more objectively judge the true maturity of this kind of technology.
