Evaluating Large Models for 3D Creation and Interactive Web Page Generation
This article evaluates the capabilities of large language models in 3D content creation and interactive web delivery. Based on a bare prompt experiment generating a voxel-style 3D miniature diorama, it explores prompt design, deliverable validation standards, and the engineering value of interactive web features. It also provides recommendations for deploying these 3D generation results on general cloud platforms.
1. Background: From Code Evaluation to 3D Creation Evaluation
For a period, evaluations of large model capabilities focused primarily on code completion, function generation, and unit test fixing. With the growing demand for 3D content generation, the focus has extended to 3D modeling, spatial reasoning, scene organization, and interactive web delivery. A typical experimental approach is to provide the tested model with a bare prompt, requiring it to generate a 3D miniature diorama with a clear theme, and deliver editable 3D project files, renderings, and an interactive web page. This type of experiment is closer to the real content production pipeline than simply looking at code snippets, as it simultaneously tests semantic understanding, geometric organization, materials and lighting, camera angles, asset export, and front-end interaction encapsulation.
The original experiment used a “Detective Agency” themed miniature diorama as the target, requiring voxel block language for the main building structure, along with editable project files and renderings. The results were quite representative: the building structure, stairs, second-floor space, ground-floor shop, and character figures were basically present, but the crucial signboard element was missing. However, the interactive web page exceeded the prompt’s requirements, providing drag-and-drop, auto-rotation, scene annotation, view focusing, asset download, image switching, and first-open guidance. This result indicates that while large models can achieve a “viewable” delivery in 3D creation, there is still a gap before reaching a production-grade delivery that is “acceptable, editable, and deployable.”
It must be emphasized that this article discusses general technical capabilities and does not revolve around any specific third-party product. Terms like “tested model,” “integration test environment,” and “quota limits” in the experiment are neutral expressions. For enterprise developers, the more valuable question is not a specific release itself, but how to evaluate 3D generation capabilities, how to design prompts, how to accept deliverables, and how to integrate generation results into a general cloud deployment pipeline.
2. Bare Prompt Experiment Design and Deliverable Breakdown
2.1 Prompts Should Cover “Theme, Style, Key Elements, Delivery Format”
The advantage of a bare prompt is that it observes the model’s default capabilities, but the disadvantage is that it easily omits key constraints. The prompt in the experiment already included the theme, voxel style, visual highlights, and deliverables, but the signboard was still missing. The reason is usually not that the model is completely unable to generate text or signboards, but that the prompt did not elevate the “signboard” to a mandatory acceptance item. A more robust prompt should explicitly list mandatory elements, optional elements, prohibited missing items, and a deliverable checklist.
A reusable template is as follows:
Please create a 3D voxel miniature diorama themed "Detective Agency in a City Block".Requirements:1. The main building structure should be prominent, including a second-floor agency, external stairs, a ground-floor shop, and a readable signboard;2. Adopt voxel block language to control polygon count and material complexity;3. Deliver an editable 3D project file, at least two renderings, and an interactive web page;4. The web page must support drag-to-rotate, auto-rotate, scene annotation, view focus, asset download, and first-open guidance;5. All key elements must be listed with names and coordinates in the scene configuration for validation.
The key to this template is writing the “readable signboard” into the mandatory elements, including interactive features in the acceptance checklist, and requiring scene configuration output. Compared to simply saying “open the file in a web page,” explicit constraints can significantly reduce delivery deviations.
2.2 Deliverable Validation Cannot Rely Solely on Renderings
The experiment delivered two renderings and an editable project file. Renderings are responsible for visual persuasion, the project file for editability, and the interactive web page for demonstration and dissemination. The acceptance criteria for these three are different:
- Renderings: Check composition, lighting, materials, theme recognizability, and whether key elements are distinguishable.
- Editable project: Check hierarchy organization, naming conventions, whether materials are replaceable, whether cameras and lighting are preserved, and whether the model can be independently edited.
- Interactive web page: Check loading performance, interaction response, annotation accuracy, focus logic, asset download, and mobile adaptation.
If scores are given based solely on renderings, issues such as chaotic hierarchy, unreadable naming, and lost materials in the project file can easily be ignored. Production-grade validation should include “editability” and “maintainability” in the scoring criteria.
2.3 Missing Key Elements Reveal Semantic Salience Issues
The missing signboard is the most noteworthy deviation in this experiment. For the model, a signboard is both a geometric body and a semantic anchor; it is not just a block on the wall, but an important part of theme recognition. Its absence indicates that the model’s understanding of “visual highlights” may stop at the building’s outline, without incorporating text identifiers, positions, and readability into the generation constraints. Engineering-wise, this can be compensated for through a “mandatory element checklist + coordinate annotation + validation script”:
{"scene": "Detective Agency Miniature Diorama","required_landmarks": [{"name": "Main Building", "must_exist": true},{"name": "Readable Signboard", "must_exist": true, "position": "Second-floor exterior wall"},{"name": "External Stairs", "must_exist": true},{"name": "Ground-floor Shop", "must_exist": true},{"name": "Character Figure", "must_exist": false}],"deliverables": ["editable_3d_project", "render_png", "interactive_html"]}
The styling deviation of the character figure also indicates that fine-grained geometry and style consistency remain difficult. Issues such as abstract hair and disproportionate scaling may be amplified in the voxel style. During validation, a distinction should be made between “theme recognizable” and “high-fidelity details” to avoid applying the same standard to all elements.
3. The “Beyond Expectations” Capabilities and Engineering Value of Interactive Web Pages
The most valuable observation in the experiment was that the prompt only required the web page to open the file, but the delivered web page supplemented drag-and-drop, auto-rotation, scene annotation, view focus, asset download, image switching, and first-open guidance. These features are not mere showing off, but general capabilities for 3D content demonstration.
3.1 Scene Annotation and View Focus Suit Visualization Projects
Scene annotation maps key objects in 3D space to readable labels, commonly implemented by maintaining object names, 3D coordinates, and screen overlays. View focus quickly brings the user to a specified object through interpolation of camera position and target point. For scenarios like real estate sandboxes, city modeling, and campus visualization, these capabilities can significantly reduce comprehension costs. During validation, check whether the annotations follow the perspective, whether they are occluded, and whether the global view can be returned after focusing.
focus(scene, target):camera.position = interpolate(camera.position, target.viewPosition)camera.lookAt(target.center)overlay.show(target.annotation)
3.2 Asset Download and Image Switching Enhance Delivery Completeness
Providing model downloads and rendering switches within the web page upgrades the deliverable from a “demo page” to an “asset entry point.” Engineering-wise, it should be confirmed that the download links point to static resources, the files are complete, naming is clear, and access logs are recorded. Image switching should avoid loading too much on the first screen; lazy loading and thumbnail previews can be used.
3.3 First-Open Guidance Lowers the Usage Barrier
Displaying drag-and-drop instructions on the first open is a classic onboarding design. In implementation, local state can be used to mark whether it has been displayed to avoid repeated interruptions. For 3D interactive pages, first-open guidance can reduce the misunderstanding of “the page is static” and improve the success rate of demonstrations.
4. From Generation to Deployment: General Cloud Landing Recommendations
For 3D generation results to enter enterprise workflows, asset hosting, page deployment, logging, and monitoring are typically required. A combination of general cloud capabilities can be used: object storage for hosting models and renderings, container platforms or static hosting for deploying interactive pages, log services for recording interaction events, and monitoring alerts for observing availability. In general cloud environments like Baidu AI Cloud, these categories can be combined as needed, but the specific selection should be based on the team’s existing tech stack and compliance requirements.
The recommended validation checklist is as follows:
- Are the theme and mandatory elements complete, and are semantic anchors like signboards readable?
- Are the project file hierarchy, naming, materials, camera, and lighting editable?
- Do the web page’s first-screen loading time, interaction frame rate, and mobile adaptation meet standards?
- Are scene annotation and view focus accurate, returnable, and free from occlusion?
- Are asset download, image switching, and first-open guidance functional?
- Are quotas and costs controllable, and is there a queue and retry mechanism for batch generation?
- Are generated assets incorporated into version control and object storage lifecycle policies?
5. Conclusion
This bare prompt experiment demonstrates that large models already possess usable structural organization and web encapsulation capabilities for 3D miniature diorama generation, with interactive web pages even compensating for features not explicitly requested. However, issues such as missing key elements, character detail deviations, and insufficient editability validation still remind developers: 3D generation cannot rely solely on renderings; a complete pipeline of “prompt constraints—deliverable validation—engineering deployment” must be established. By writing mandatory elements, scene configurations, interactive features, and asset checklists into prompts and validation scripts, and combining them with general cloud hosting, logging, and monitoring capabilities, a demonstrative generation can be transformed into a reusable 3D content production practice.
