Evaluating General Large Models for 3D Creation and Web Delivery Capabilities
This article evaluates the capabilities of general large language models in 3D creation, spatial reasoning, and interactive web delivery. Using a voxel diorama of a detective agency as a practical case study, it analyzes model performance, highlighting challenges such as the omission of key semantic elements and uncertainties in engineering editability. The article provides an engineering framework featuring structured prompts, a JSON delivery contract, and automated validation to transform model-generated 3D assets into production-ready deliverables.
Recently, the capabilities of general large language models have expanded from simple text Q&A and code completion to 3D creation, spatial reasoning, and the generation of interactive deliverables. Unlike “generating a piece of code” or “outputting an image,” 3D creation tasks simultaneously test a model’s comprehensive understanding of spatial structures, object relationships, material styles, engineering file organization, and web frontend interaction. These tasks are closer to real-world development scenarios: users need not just a final result, but editable, verifiable, and re-developable assets and pages.
This article centers on a 3D creation practice: using a “detective agency themed voxel diorama” as the task, we employed a bare prompt approach without attaching additional skills or tools. We observed the 3D engineering files, render images, and web interactive pages delivered by the model to extract reusable evaluation methods and engineering recommendations. Please note that this article discusses general capabilities and engineering practices and does not promote or test any specific product.
1. Testing Goals: From “Can Generate” to “Deliverable”
The motivation for the original test is clear: when a model claims to have 3D creation and spatial reasoning capabilities, looking at promotional materials is insufficient to judge actual usability. A more reasonable approach is to design a task with a clear theme, key elements, and delivery requirements, allowing the model to perform freely under minimal constraints, and then assessing the completeness of the delivery through an acceptance checklist.
Key requirements for this task included: presenting the detective agency theme using voxel blocks; highlighting the main building structure; delivering editable 3D engineering files and render images; and additionally requiring an interactive web display that can be opened in a browser. This design simultaneously covers three types of capabilities: semantic understanding and theme restoration, 3D asset construction, and web interactive page generation.
From an evaluation perspective, we cannot just look at whether it “looks right.” We must also check if key elements are missing, if the engineering files are editable, if object naming is clear, if the render images are usable, and if the web page has basic interactivity. For enterprise users, the latter is often more critical because deliverables must enter design, presentation, or visualization project workflows rather than remaining as demo screenshots.
2. Bare Prompt Testing: Free Expression and Critical Omissions
We tested using a bare prompt without additional skills or tools to observe the model’s default delivery strategy under natural language instructions. The initial prompt requested the creation of a voxel-style diorama, delivering an editable 3D engineering project and render images. The model ultimately delivered two render images and an editable 3D engineering file. The overall restoration was acceptable: structures such as the main building, the second-floor studio, the first-floor shop, and the external stairs were basically present, and a character model was placed at the entrance.
However, the problem was also obvious: the most eye-catching signboard element of the theme was missing. For a “detective agency” theme, the signboard is not just decoration; it is a point of semantic recognition. Without it, the model built the building but was incomplete in terms of key narrative elements. Therefore, this delivery cannot be judged as fully passing the acceptance criteria. This phenomenon indicates that during free generation, the model tends to prioritize “building morphology” and “style language” while ignoring the most critical identifying objects in the user’s mind.
This also serves as a reminder for developers: prompts should not only specify the theme and style but also write “must-exist key elements” as hard constraints. Otherwise, the model might fill the scene with reasonable but non-critical elements, resulting in an outcome that “looks good but is unusable.”
3. Web Interactive Pages After Supplementary Prompts: The Surprises
After supplementing the instruction to “open the deliverable in a web page,” the page generated by the model brought several surprises. The page not only supported model drag-to-rotate and auto-rotate but also included scene annotations, camera focus, model download, image switching, and a drag guide on first open. For projects like smart city sandboxes, campus displays, and architectural visualization, these features have direct reference value.
Scene annotations and camera focus are particularly noteworthy. They transform the “3D model” from a static asset into an explainable presentation interface: users can click annotations to locate objects like the main building, studio, and shop, or use focus buttons to quickly switch to preset views. Model download and image switching complete the asset distribution capability. The tone, lighting, and filter processing of the render images were also quite complete, indicating that the model has a certain stability in visual presentation.
However, this capability carries uncertainty. Although the page was feature-rich, partly because the model proactively expanded the concept of “web display,” if users need strict control over page structure, component libraries, performance budgets, and accessibility standards, natural language generation alone is insufficient. For production environments, page generation must still be decomposed into modules—page skeleton, interaction logic, asset loading, annotation configuration, and download interfaces—supplemented by manual review.
4. Capability Boundaries: 3D Creation Still Requires an Acceptance Loop
From this practice, we can summarize several capability boundaries. First, key semantic elements may be omitted. A model can generate a building main body but may not understand the importance of narrative elements like signboards, logos, and character relationships. Second, engineering editability needs verification. Delivering an engineering file does not mean the object hierarchy, materials, units, and naming meet secondary development requirements. Third, web interactive pages are random. The model may proactively add features but might also omit engineering details like resource loading, error handling, and mobile adaptation. Fourth, resource quotas and invocation costs are practical pain points. 3D and page generation often require multiple rounds of interaction, and insufficient quotas limit large-scale scene creation, multi-round iterations, and comparative testing.
Therefore, evaluation should not stop at “first-generation results.” A more reasonable approach is to establish an acceptance loop: first define the delivery contract, then generate, then automatically validate, and finally manually review. Below is an example of a general delivery contract, where fields can be adjusted based on the project:
{"task": "detective_office_voxel_diorama","deliverables": {"editable_project": true,"render_images": ["front.png", "perspective.png"],"web_viewer": {"orbit_controls": true,"auto_rotate": true,"annotations": true,"focus_views": true,"download_assets": true,"image_switch": true}},"acceptance": {"key_signage": "must_exist","building_structure": "complete","style": "voxel","editable": true,"object_naming": "clear"}}
The corresponding prompt should also be structured, separating “must-haves” from “optionals.” For example: Please create a voxel-style diorama themed around a detective agency using a general 3D modeling process; it must include the main building structure, a second-floor studio, a first-floor shop, external stairs, and prominent signboard text; character models are optional; deliver an editable 3D engineering project, at least two render images, and a web interactive page; the web page must support drag-to-rotate, auto-rotate, scene annotations, camera focus, asset download, and image switching; all object naming should be clear, and materials and dimensions should use uniform units.
5. Engineering Recommendations: Turning “Demo Effects” into “Reproducible Processes”
For developers and enterprise users, it is recommended to break down 3D creation tasks into reproducible processes. The first step is to define the delivery contract, clarifying file types, object lists, key elements, interactive features, and acceptance criteria. The second step is to use phased prompting: first generate the building structure and key identifiers, then generate materials and styles, and finally generate the web page to avoid one-time instruction overload. The third step is to establish automated validation scripts to check whether engineering files can be opened, key objects exist, the number of images is complete, and page resources can be loaded. The fourth step is to introduce manual review, focusing on semantic restoration, visual style, and secondary editing costs.
At the infrastructure level, deliverables can be placed in object storage, log services can be used to record generation rounds and failure reasons, and monitoring alerts can track page load failure rates and resource quota consumption. For multi-round iteration projects, prompt versions, model output versions, and acceptance results should also be saved to form a traceable evaluation record. This not only reduces the cost of repeated trial and error but also provides a basis for subsequent model selection.
6. Conclusion
This general large model 3D creation practice demonstrates that models already exhibit strong usability in building structure restoration, voxel style expression, render image generation, and web interactive page expansion. They can even proactively add scene annotations, camera focus, and download capabilities under simple prompts. However, the omission of key elements, uncertainty in engineering editability, randomness in page interaction, and resource quota costs still require developers to establish a strict acceptance loop.
A more reasonable attitude is to view 3D creation as the starting point for generating “deliverables” rather than the endpoint. Through structured prompts, delivery contracts, automated validation, and manual review mechanisms, the model’s free expression can be constrained within an engineering-usable range. Only when key elements, engineering files, rendered assets, and interactive pages all pass acceptance does 3D creation capability truly transition from a demo effect to a productive force.
