Evaluating LLMs' 3D Creation Ability: From Bare Prompts to Interactive Web Pages
As large model evaluations move beyond text and code, this article assesses 3D creation using a detective office diorama in a voxel style, built from a bare prompt. It covers the dimensions to evaluate, the deliverables received, and a reusable checklist. Readers will see why 3D creation tests a model's spatial reasoning, procedural scripting and engineering completeness, why delivery formats depend heavily on prompt wording, which key element the model failed to generate, and what interactive web features the model added beyond the prompt. They will also get guidance on verifying that project files can be edited and assets are complete.
Introduction: Why 3D Creation Is Becoming a New Focus for Large Model Evaluation
Over the past two years, the focus of large model evaluation has gradually shifted from text understanding and code generation toward multimodality and spatial intelligence. One trend stands out: comparing code correctness rates alone is no longer enough to capture the differences between models. The industry is beginning to measure real capability through deliverables you can see and rotate, such as 3D models and interactive web pages.
3D creation is an ideal benchmark for three reasons. It tests geometric and spatial reasoning, since the model must understand how objects relate in position, structure and proportion. It also tests procedural expression, because much 3D generation boils down to writing scripts, organizing scene graphs and controlling materials and cameras. Finally, it tests engineering completeness: can the model deliver output that is editable, reusable and open to further development?
This article follows a real 3D scene generation practice and distills it into a reusable evaluation methodology that developers can apply when assessing this kind of capability.
1. Evaluation Design: Define What You Are Testing First
Before designing an evaluation, you must break the capability down into observable dimensions. Otherwise, it is easy to fall back on subjective judgments such as “looks about right, let’s give it a score.” A serious 3D generation evaluation should cover at least the following four levels:
- Thematic fidelity: Given a theme, can the model capture the key visual elements, including the main subject, its defining structures and supporting props?
- Geometric and spatial correctness: Is the structure coherent? Are the proportions reasonable? Do objects intersect or float in mid-air?
- Deliverable completeness: Does the output include editable project files, rendered images and, if needed, an interactive showcase page?
- Usability and engineering quality: Is the output well organized? Can others take over and keep developing from it?
Once the dimensions are clear, the evaluation has a solid basis to stand on.
2. Bare Prompts: The First Step in Controlling Variables
To rule out interference from attached skills or plugins, evaluations should rely as much as possible on bare prompts, meaning prompts used without any external add-ons. Results obtained this way better reflect a model’s core, out-of-the-box ability.
In this practice, the prompt provided only three things:
- A theme: a miniature scene inspired by a detective office.
- A style constraint: a voxel, block-based visual language.
- The desired deliverables: an editable 3D project file and rendered images.
This minimally constrained design keeps the variables clean. It also shows what the model does by default when information is missing.
One key observation emerged. When the prompt did not mention an interactive web page, the model delivered only a static model and rendered images. An interactive page was produced only after the prompt was extended with a request to open the file in a web browser. In other words, the form of the deliverables depends heavily on explicit guidance in the prompt.
Evaluators should therefore treat “does the prompt fully express the intent?” as an important item to watch.
3. Deliverable Review: Solid Fidelity, but Details Make the Difference
In terms of geometric fidelity, the model’s output was acceptable. The core structures expected in the reference scene were largely present:
- A staircase leading up to the second floor
- An office space on the second floor
- A dining business on the first floor
- Two iconic characters placed at the entrance
The problem lay in the most critical detail, though. The most recognizable sign in the scene was not generated, even though it is the visual focal point of this particular theme. The verdict is therefore nuanced: overall fidelity was decent, but the missing key identifying element kept the output from earning full marks.
The lesson for evaluators is that a 3D generation evaluation cannot look only at how closely the result resembles the reference. It must also check whether the elements that should be there are actually there. It is worth assigning a dedicated weight to “key identifying elements” in the score sheet. Otherwise, a model may mask missing core elements with a mass of irrelevant details.
Abstracted details such as the characters’ hairstyles and proportions should be recorded as well. They provide important clues about the model’s aesthetic judgment and grasp of human anatomy.
4. The Interactive Web Page: Value Beyond the Prompt
The real surprise came from the interactive web page generated after the prompt was extended. It went beyond the explicit request and proactively added several practical features:
- Basic interaction: standard 3D preview controls such as drag-to-rotate and auto-rotation.
- Scene annotations: text labels on key structures. This capability is especially useful for real estate scale models, city modeling, smart visualization and similar use cases.
- View focusing: a sidebar lists several scene elements, and clicking one moves the camera to focus on the corresponding part, making it easy to locate it quickly.
- Asset integration: links to download the model and a switchable gallery of multiple rendered images, all brought together on the same page.
- First-time onboarding: a hint on drag controls when the page is first opened, lowering the learning curve for new users.
From an engineering perspective, this ability to generate a complete, interactive page on the spot shifts much of the work into the generation stage. Previously, this meant frontend developers manually building a 3D viewer. When evaluating 3D generation capabilities, the completeness of the interactive page should therefore count as a separate bonus criterion.
5. Rendering Quality and Engineering Conventions
Beyond functionality, quality at the rendering and engineering levels also deserves attention. In this run, the output handled render filters and lighting well. That suggests the model has a reasonable grasp of default material and lighting settings. Evaluators should still check the following:
- Editability: Can the project files be opened and modified normally in mainstream 3D software, rather than being read-only or corrupted?
- Structural conventions: Are object names, layer organization and material names clear enough to support further development?
- Resource self-containment: Are any external textures or assets missing? If so, materials may fail to load when the project is opened.
These engineering details often determine whether generated content can actually enter a production workflow.
6. A Reusable 3D Generation Evaluation Checklist
Drawing on this practice, here is a 3D generation evaluation checklist that can be put to use right away:
- Control the variables: Use bare prompts first to avoid interference from plugins or skills.
- Score by dimension: Rate thematic fidelity, geometric correctness, key elements, the interactive page and engineering conventions separately.
- Prioritize what matters: Give a dedicated weight to the most recognizable key elements so that irrelevant details do not dilute the focus.
- Verify editability: Confirm that project files can be opened and modified and that all resources are complete.
- Observe added value: Check whether features of the interactive web page, such as annotations, focusing and downloads, go beyond expectations.
- Note resource limits: Watch for platform limits such as usage quotas and daily request caps, and pace the evaluation accordingly.
Conclusion
What looks like a simple “generate a miniature scene” test actually brings together spatial reasoning, script generation, engineering delivery and interactive design. The value of an evaluation does not lie in giving a model an overall score. It lies in presenting the boundaries of a model’s capability objectively, through clear dimensions and rigorous acceptance criteria: where it meets the bar, where it falls short and where it exceeds expectations.
For developers, building this methodology into a repeatable process is what makes it possible to reach judgments that are comparable, reproducible and actionable in a fast-moving model ecosystem.
