Evaluating Large Models on 3D Creation and Interactive Web Pages: A Practical Framework
As multimodal and structured generation advance, evaluation of large models is shifting from single-answer quality toward the ability to deliver complex artifacts. This article presents a practical framework for evaluating 3D creation and interactive web generation: a detective-office voxel diorama task, a multi-dimensional scoring system, common failure patterns, prompt-engineering techniques, and guidance for moving model output into real production pipelines.
1. From Generating Text to Delivering Space: Why 3D Became a New Evaluation Focus
For a long time, evaluating large models meant looking at text comprehension, code completion, and reasoning in question answering. More recently, as generative models have improved in multimodal output, structured generation, and interactive front-end pages, the focus of evaluation is shifting from “quality of a single answer” toward “the ability to generate complex deliverables.” Within that shift, 3D modeling, spatial reasoning, scene assembly, and interactive web generation have become important dimensions for observing a model’s overall capability.
There are clear technical reasons behind this change. 3D creation is not simple text generation; it requires the model to understand information across several layers at once:
- Spatial relationships: The up/down, left/right, front/back, floor, entrance, stairway, and passage relationships between objects must make sense.
- Semantic completeness: A themed scene must not only “look the part” but also contain the key elements the theme requires.
- Geometric consistency: Voxels, meshes, and model components need to keep basic proportions, avoiding clipping, floating parts, and distorted scale.
- Materials and rendering: Color, lighting, filters, background, and viewpoint all affect the final visual result.
- Interaction: If an interactive web page is required, this also involves front-end logic such as dragging, rotating, zooming, annotation, view switching, and asset download.
- Engineering maintainability: Whether the output can be further edited, layered, imported, and have its assets replaced determines whether it can enter a real project workflow.
So evaluating a newly released 3D generation capability cannot rest on whether a single rendered image looks good. It has to examine whether the model can complete the full loop from “understanding requirements” to “spatial construction” to “interactive delivery.” A model released late at night drew attention not because of its release time, but because developers began to ask: Can this kind of capability enter a real production workflow? Does it have reproducible acceptance criteria? Can it support actual scenarios such as city sand tables, virtual exhibition halls, digital asset libraries, and training simulations?
2. Evaluation Task Design: Testing Spatial and Semantic Capability with a “Detective Office” Voxel Diorama
To test a model’s 3D creation ability, you can use a task that is moderately structured, has clearly defined elements, and makes differences easy to observe: generating a detective-office-themed 3D voxel miniature scene.
The reason for choosing this kind of task is that it contains several observable evaluation points at once:
- Main building: Does the office building have a clear structure?
- Spatial layering: Are there spatial relationships such as a first floor, second floor, entrance, stairs, windows, and porch?
- Thematic elements: Are there recognizable symbols such as a coffee shop, an office sign, characters, and signage?
- Voxel style: Does it use a block-based language rather than realistic modeling?
- Deliverable completeness: Can it output an editable model, rendered images, and an interactive page?
- Detail fidelity: Are characters, signage, stairs, and scene annotations natural and reasonable?
This kind of task is more discriminating than simply generating a “house.” The model must not only generate geometry but also understand the theme, spatial layout, visual style, and interactive expression. It can expose the boundaries of a model’s capability in complex scenes: which parts can be generated automatically, which parts are easily missed, which parts only “look present,” and which parts are genuinely usable in a project.
2.1 The Value of Bare-Prompt Testing
“Bare prompt” here means completing the delivery using only natural-language prompts, with no extra skills, templates, plugins, external knowledge bases, or human-written scripts. The value of this test method is that it lets you observe:
- The model’s own spatial understanding;
- Whether the model can proactively fill in the types of deliverables;
- The model’s completion strategy when requirements are incomplete;
- Whether the output has real engineering usability.
In real projects, developers rarely start with a fully detailed requirements document. Often, a user gives just a theme, a style, and a delivery goal. Bare-prompt testing is therefore closer to real interaction scenarios.
2.2 A Structured Prompt Example
While bare prompts can reveal a model’s ceiling, structured prompts help produce stable output in engineering practice. You can break requirements into modules such as theme, style, layout, required elements, interactive features, and delivery format. Below is an illustrative structure:
{"task": "create_voxel_micro_scene","theme": "detective_office","style": "voxel","composition": {"building": "two_story_townhouse","floor_1": ["entrance", "coffee_shop", "stairs"],"floor_2": ["office", "signboard", "windows"],"characters": ["two_figures_near_entrance"]},"required_elements": ["signboard","stairs","coffee_shop","office_room","two_characters"],"deliverables": ["editable_3d_project","rendered_images","interactive_web_page"],"interactive_features": ["drag_rotate","auto_rotate","scene_annotations","focus_points","image_switching","model_download"],"constraints": {"no_third_party_ip": true,"keep_clear_geometry": true,"avoid_abstract_facial_features": true}}
The benefit of this kind of structured input is that it turns vague requirements into acceptance conditions. The model is no longer just “drawing a scene that looks like a detective office”; it must organize space, configure elements, and produce interactive output around clearly defined objects.
3. Evaluation Dimensions: Don’t Look Only at Rendered Images—Look at the Full Delivery Loop
For 3D generation tasks, it is advisable to build a multi-dimensional scoring system. Looking only at rendered images invites misjudgment, because a beautiful image can hide missing structure, unusable interaction, and non-editable assets. A more reasonable evaluation should cover the following aspects.
3.1 Semantic Fidelity
Semantic fidelity concerns whether the model understands the theme and generates the key elements it should have. Taking the detective-office voxel diorama as an example, it should at minimum contain the office building, entrance, stairs, interior spaces, a themed signboard, and character figures. If the most critical visual marker is missing, even a complete rest of the scene cannot earn a high score.
How to check:
- List the elements the theme must include;
- Compare whether each item is present in the model’s output;
- Mark the severity of each missing item;
- Determine whether a miss is due to an unclear prompt or the model’s own insufficient understanding.
3.2 Spatial Structure Reasonableness
The most basic requirement for a 3D scene is that spatial relationships hold. Floors, stairs, doors and windows, entrances, and interior layouts must not show obvious violations of common sense. For example, stairs should connect the upper and lower floors, doors should not float, the building facade should have reasonable openings, and interior spaces should not conflict with external structure.
How to check:
- Observe from front, side, top-down, and interior views;
- Check for floating parts, clipping, and abnormal occlusion;
- Check whether character proportions match the building scale;
- Check whether interactive viewpoints can display key areas without obstruction.
3.3 Geometry Quality and Voxel Style
If the task requires a voxel style, the output should express a block-based language rather than suddenly switching to a realistic model. The point of voxel style is not precision but a sense of order: whether the blocks are neat, whether proportions stay stable, whether details are overly complex, and whether components are easy to extend.
How to check:
- Observe whether the model keeps a consistent block-face language;
- Check whether components can be edited independently;
- Look for large amounts of uncontrollable mesh;
- Assess the feasibility of later replacing materials and adding modules.
3.4 Rendering and Visual Completeness
Rendered images are the developer’s first impression of output quality. But rendering should not be separated from structural quality. Lighting, shadows, background, filters, and viewpoint all affect perception. A scene whose structure is chaotic cannot claim to meet the bar just because filters polish it.
How to check:
- Are multiple-angle renders provided?
- Does the lighting make spatial relationships clearer?
- Is the subject prominent?
- Are there over-blurring, color distortion, or texture stretching?
- Do the renders match the 3D model?
3.5 Interactive Web Page Capability
If the model further generates an interactive web page, the evaluation focus expands from “static assets” to a “lightweight application.” A good interactive page should let an ordinary user quickly understand the scene and browse it without a development background.
Common interactive capabilities include:
- Mouse drag to rotate;
- Auto-rotation display;
- Zoom and pan;
- Scene annotation;
- Focusing on key elements;
- Rendered-image switching;
- Model download;
- Onboarding guidance on first entry.
These capabilities are not showing off; they address basic needs in real projects. For example, in city sand tables, property showcases, park guides, and exhibition descriptions, scene annotation and viewpoint focus can directly improve how efficiently information is conveyed.
3.6 Engineering Editability
Many 3D generation tasks overlook engineering editability. If what developers receive is only an image or a closed file, maintenance costs later will be high. A deliverable that is genuinely usable in a project should support continued editing, importing, splitting, material replacement, and camera and animation adjustments.
How to check:
- Does the project file open?
- Is the model layered?
- Are materials modifiable?
- Are cameras and lights independent?
- Can it export to common formats?
- Is it easy to plug into a later visualization workflow?
3.7 Performance and Resource Consumption
Delivering 3D on the web also means considering performance. If a model has too many polygons, oversized textures, or disordered loading logic, it will hurt the browser experience. For enterprise projects, a performance budget is especially important.
Observable metrics include:
- First-load time of the page;
- Size of the 3D model;
- Number and resolution of rendered images;
- Whether progressive loading is supported;
- Whether auto-rotation can be turned off;
- Whether the page can still be browsed basically on low-performance devices.
3.8 Safety and Compliance Risks
3D-generated content also needs to account for risks involving copyright, trademarks, character likenesses, and scene identifiers. If a theme involves well-known stories, character figures, or brand signage, protected identifiers should not be generated directly. For commercial projects, it is better to use original themes or clearly licensed asset libraries.
Safety acceptance recommendations:
- Do not generate recognizable third-party trademarks;
- Do not use unauthorized story characters;
- Make signage, text, and character designs original;
- Keep review records for generated content;
- Manually review high-risk assets.
4. Typical Observations: Clear Progress, but Semantic Omissions Remain
From a typical detective-office voxel diorama test, model output usually shows several characteristics.
4.1 The Main Building Is Fairly Well Reconstructed
Models often manage to generate basic structures such as multi-story buildings, entrances, stairs, shop spaces, and office areas. This shows they have some ability to understand a “building theme” and can turn an “office” into a visualizable space rather than generating only isolated objects.
For example, elements such as side stairs, a second-floor office, a first-floor coffee area, and characters at the entrance are usually recognizable. This suggests the model already has some scene-organization ability, able to combine multiple spatial modules into a relatively complete miniature scene.
4.2 Interactive Web Pages Often Exceed the Original Request
When only a 3D model and rendered images are requested, the model may also produce an interactive web page. This is worth noting: it shows the model can not only generate assets but also proactively fill in front-end presentation logic.
Commonly added features include:
- Drag to rotate the model;
- Auto-rotation;
- Scene annotation;
- Viewpoint focus;
- Model download;
- Rendered-image switching;
- First-load guidance.
If controlled well, these capabilities significantly improve the usability of the deliverable. For instance, in a building sand-table showcase, auto-rotation helps users quickly grasp the building’s appearance, scene annotation can explain the function of different areas, and viewpoint focus can direct attention to key assets.
4.3 Key Markers Are Easily Missing
A typical problem is that the most recognizable element in the scene may be missing. In the detective-office theme, the signboard on the wall or above the doorway is a key visual symbol. If the model does not generate the signboard, or if the signboard’s content is blurry and unreadable, the theme’s recognizability drops noticeably.
This shows the model still has biases in semantic understanding: it can understand the “office” as a building space, but may not reliably understand the “office signboard” as a required visual object. For developers, this means key elements must be stated explicitly in the prompt, with position and acceptance criteria given as far as possible.
4.4 Character Details Tend to Become Abstract
If the task includes characters, the model may generate figures at the entrance, staff, or visitors, but their features are not necessarily stable. Hairstyles, facial features, clothing, and proportions may become abstract, symbolic, or even unrealistically pieced together.
This shows that character generation remains a difficulty in 3D scene generation. Characters need not only a form but also proportion, posture, and an interactive relationship with the scene. If a character exists only as decoration, the model may give a simplified version; if a character must carry a narrative function, the demands on the model are higher.
4.5 Feature Completion Reflects Proactive Design
In some tests, after generating a page, the model proactively adds features such as viewpoint focus, scene annotation, a download button, and image switching. These features may not have been stated explicitly in the original prompt, yet they do improve the completeness of the page.
This suggests that when evaluating a model’s 3D capability, we should not only ask “did it generate a model?” but also ask:
- Does the model understand how the deliverable will be used?
- Does it consider how users will view it?
- Does it consider how developers will download it?
- Does it consider how non-expert users will interact?
- Does it consider page loading and interaction guidance?
These questions are closer to real projects than simply generating a static model.
5. A Sample Scoring Table: Turning Evaluation from Subjective Feeling into Engineering Judgment
To make evaluation more reproducible, you can design a scoring table. Different projects can adjust weights according to their goals, but the basic structure should stay consistent.
| Dimension | Suggested Weight | Key Checkpoints | Common Deductions |
|---|---|---|---|
| Semantic completeness | 20% | Are thematic elements complete? | Missing signage, characters, key areas |
| Spatial structure | 20% | Are floors, stairs, entrances reasonable? | Floating parts, clipping, imbalanced proportions |
| Geometry quality | 15% | Is the model clean and editable? | Chaotic meshes, non-separable components |
| Rendering | 10% | Lighting, viewpoint, clarity | Blurry images, weak subject |
| Interaction | 15% | Drag, annotation, focus, download | Laggy page, missing interaction |
| Engineering usability | 10% | Is it easy to import and re-edit? | Disorganized file structure |
| Performance & compatibility | 5% | Load speed, page stability | Oversized assets, browser errors |
| Compliance & safety | 5% | Does it avoid infringement and sensitive content? | Using unauthorized characters or marks |
The key to this scoring table is breaking “looks good” into verifiable metrics. Developers can adjust weights by project type. For example:
- Digital exhibition projects: higher weight on semantic completeness and interaction;
- City sand-table projects: higher weight on spatial structure and performance;
- Early game or animation work: higher weight on geometry quality and editability;
- Brand showcase projects: higher weight on compliance/safety and rendering.
6. From Evaluation to Implementation: How 3D Generation Enters Real Projects
The value of model-generated 3D scenes is not only in showing a technology trend, but more importantly in entering real business workflows. For developers or enterprise users, the real questions include: Can it generate in batches? Can it be re-edited? Can it be deployed to the web? Can it integrate with existing systems? Can costs and risks be controlled?
6.1 Suited to Digital Sand Tables and Spatial Displays
In scenarios such as property showcases, park guides, city sand tables, and exhibition displays, model-generated 3D assets can reduce the cost of early-stage prototyping. Voxel-style, low-poly, and schematic styles are especially suited to expressing spatial relationships quickly.
These scenarios usually do not need extreme realism; they care more about:
- Whether building outlines are clear;
- Whether area divisions are distinct;
- Whether key points can be clicked;
- Whether viewpoint switching is smooth;
- Whether annotation text is readable.
So when model-generated results are paired with an interactive web page, their value increases noticeably.
6.2 Building an Asset Library and Standardized Workflow
If you want to use 3D generation for long-term projects, you cannot rely on one-off prompts. A more reasonable approach is to build a standardized asset library and task templates.
For example, assets can be divided into:
- Building modules;
- Roads and ground;
- Vegetation and landscape;
- Characters and traffic;
- Functional markers;
- Lights and cameras;
- Interaction hotspots;
- Text annotations.
For each module, define naming conventions, coordinate standards, material standards, and export formats. This way, model-generated content can enter the project pipeline faster rather than being rebuilt every time.
6.3 Completing Delivery with General Cloud Capabilities
In enterprise applications, 3D generation is often not an isolated step but needs to combine with content storage, distribution, monitoring, and deployment. General cloud product categories can provide the support, for example:
- Object storage: to store model files, rendered images, and texture assets;
- Content delivery: to speed up access to static resources;
- Container platforms: to deploy interactive 3D web pages;
- Log services: to record page loads and interaction behavior;
- Monitoring and alerting: to detect load failures and resource anomalies;
- Message queues: to handle asynchronous generation tasks;
- Identity and access management: to control the scope of asset access.
This turns 3D generation from a one-time “model output” into a sustainable content-production capability.
6.4 Human-AI Collaboration Remains Key
Currently, model-generated 3D scenes are better suited as an efficient first-draft tool rather than a full replacement for professional modeling. For key projects, a human-AI collaboration mode is recommended:
- The model generates the initial scene;
- A human checks spatial structure and semantic completeness;
- Missing elements are addressed with additional prompts;
- Parts of the model are regenerated locally;
- Assets are edited in professional software;
- Finally, they are published to the interactive page and accepted.
This workflow preserves the efficiency of model generation while ensuring delivery quality.
7. Prompt Optimization Tips: Reducing Missing Key Elements
Many 3D generation problems are not because the model cannot do it, but because the prompt does not state the key constraints clearly. To improve output stability, you can optimize prompts from the following angles.
7.1 Specify the Deliverable Type
Do not just say “make a scene”; state whether you need a model, images, or an interactive web page. For example:
- “Please output an editable 3D project file.”
- “Please provide at least two rendered images from different angles.”
- “Please generate a page that can be opened and interacted with in the browser.”
7.2 List Required Elements
For a themed scene, list the elements that must not be missing. For example:
- The main office building;
- The second-floor office area;
- The first-floor coffee space;
- Side stairs;
- Characters at the entrance;
- Signage on the wall or above the doorway.
This makes it easier for the model to understand “what is core versus optional.”
7.3 Provide the Spatial Layout
If the model understands the relationship between floors, entrances, and stairs unstably, you can give a simplified layout in the prompt:
The first floor contains an entrance, a coffee space, and stairs leading to the second floor;The second floor contains the office workspace, with a themed signboard on the facade;Two figures stand at the front of the building, with proportions matched to the door height.
This kind of description is more reliable than simply saying “make a detective office building.”
7.4 Specify Interaction Requirements
If you want a web page, state the interaction goals:
- Support drag to rotate;
- Support an auto-rotation toggle;
- Support scene-element annotation;
- Support focusing on key viewpoints;
- Support model download;
- Support switching rendered images;
- Show interaction guidance on first open.
7.5 Define Style Boundaries
If you require a voxel style, emphasize:
- Use a block-based language;
- Avoid realistic materials;
- Keep proportions simple;
- Simplify character forms moderately, but do not over-abstract them;
- Text or signage should be clear and readable.
This reduces style drift.
7.6 Add Acceptance Criteria
You can append acceptance conditions at the end of the prompt:
Acceptance requirements:1. A visible themed signboard must be included;2. The stairs must connect the upper and lower floors;3. Characters must not float;4. The web page must support dragging and zooming;5. If complex characters cannot be generated, prioritize correct building structure.
This approach significantly improves how evaluable the output is.
8. Common Problems and Solution Approaches
In real testing, 3D generation results commonly fall into a few categories, and different strategies can be applied to each.
8.1 Missing Key Markers
If the model omits the signboard, text, or logo, raise its priority in the prompt and state its position clearly. For example: “A themed signboard must be placed at the center of the second-floor exterior wall and serve as the visual center of the scene.” If the model is unstable at generating text, switch to an abstract marker graphic and add the text manually.
8.2 Chaotic Spatial Relationships
If stairs, doors, or floor connections are unreasonable, reduce task complexity and ask the model to first generate a gray-box structure, then add details step by step. You can also split the building into several local tasks—for example, generate the building shell first, then the interior layout, and finally add characters and markers.
8.3 Character Distortion
When character generation is unstable, temporarily remove complex characters and preserve the scene’s main subject first. If characters must be included, use low-poly, silhouette, or block-figure styles to lower the detail requirement.
8.4 Incomplete Web Interaction
If the model generates a web page but interaction is missing, list the interaction features explicitly. You can also ask the model to prioritize core interactions first: drag, rotate, zoom, annotation, and download.
8.5 Files That Are Hard to Edit
If the deliverable cannot be further edited, require layered naming, independent components, and replaceable materials in the prompt. For commercial projects, also require the model to output common formats to avoid lock-in.
9. Conclusion: 3D Capability Is Worth Watching, but Must Be Accepted with Engineering Rigor
Overall, the new model has shown clear progress in 3D creation and spatial reasoning. It can generate buildings, floors, stairs, indoor and outdoor spaces, characters, and interactive web pages around a theme, showing that the model is moving from “generating a piece of code” or “generating an image” toward “delivering a usable scene.”
At the same time, evaluation must stay rigorous. A model may achieve high visual completeness yet omit key semantic elements; it may generate an interactive web page yet carry risks in engineering editability, performance, or copyright compliance. So you cannot give full marks just because a rendered image looks good, nor dismiss all capability because of one small flaw.
For developers, a more reasonable attitude is:
- Treat 3D generation as a productivity tool, not a fully automated production result;
- Use structured prompts to improve stability;
- Use a multi-dimensional scoring table for acceptance;
- Bring model output into an asset library and project pipeline;
- Keep human review for key scenes;
- Use engineering methods to control cost, performance, and compliance risk.
The early-morning release itself is not important; what matters is whether the model can withstand testing on real tasks after release. Bare prompts can expose a model’s ceiling, structured prompts can improve usability, interactive web pages can expand application scenarios, and engineering acceptance determines whether it can truly enter business systems. In the future, 3D generation capability will grow ever more important, but only by solving spatial semantics, geometry quality, interaction experience, asset maintenance, and compliance/safety together can a model become a reliable production tool for developers.
