Evaluating LLMs for 3D Scene and Interactive Webpage Generation
This article explains why 3D scene and interactive webpage generation are becoming important areas for LLM evaluation, how to design clear and reproducible test tasks, and how to assess outputs across six dimensions. It also presents an illustrative weighted scoring method, observations from a detective-agency-themed test, quota and cost management practices, and recommendations for adopting these evaluations in real projects.
1. Why 3D Generation Is Becoming a Key Focus in Model Evaluation
For some time, LLM evaluation has concentrated on code completion, algorithm problems, and text reasoning. As model capabilities expand, more developers are paying attention to another class of tasks: generating editable 3D scenes, interactive webpages, and printable model files directly from natural language descriptions. These tasks place three requirements on a model.
First, the model needs to understand spatial relationships, such as how building floors connect and where furniture sits relative to walls. Second, it needs to map abstract descriptions to concrete geometric structures while maintaining a consistent style. Third, the output must be something that can be further edited and run, not merely a rendered image.
For developers, this means evaluation cannot stop at whether the final image looks good. It must also examine whether the deliverables are complete, whether the structure is usable, and whether the interaction can be implemented in practice.
2. Principles for Designing Test Tasks
An effective test first requires a clear task description so that evaluation results are not attributed to prompt ambiguity. The following principles are recommended:
- Define the theme clearly: Specify the scene type, style, scale, and core visual elements. For example: “A small two-story building with a coffee shop on the first floor, a detective agency on the second floor, and two character models at the entrance.”
- List key elements explicitly: For required identifiers such as signs, placards, or decorations in specific colors, list them separately in the prompt and specify their positions.
- Specify deliverables: State the required file types, such as project files, rendered images, interactive webpages, or model files. If deliverables are not specified, the model may make its own choices, which complicates comparison.
- Record versions: Save the full prompt, model version, test time, and quota consumption for each test to support reproduction and comparison.
If supplementary instructions are needed during a test, record the results before and after the supplement separately. For example, if the first prompt did not mention webpage delivery, a later supplement may change the output structure, so the two results should not be compared directly.
3. Breaking Down the Evaluation Dimensions
It is recommended to split evaluation into six dimensions and provide observable criteria for each.
(1) Scene completeness: Check whether the main structure is complete. For a building, for example, verify whether the floors, stairs, entrance, and functional zones all exist.
(2) Key element recognition: Check whether elements explicitly required in the prompt appear. This dimension carries the highest weight because it directly reflects whether the model understood the task requirements. If a prominent identifying element is missing, points should be deducted even if other parts are restored well.
(3) Spatial structure and proportion: Check whether floor heights, door and window positions, and furniture placement are reasonable, and whether there are issues such as model intersections or floating objects.
(4) Detail and character performance: Check whether the shapes of characters or small objects conform to common sense. Distortion in details such as hairstyle or clothing directly affects visual perception and credibility.
(5) Interactivity: If a webpage is delivered, check whether the model supports dragging, rotation, and zooming, whether it provides initial operation guidance, and whether it offers annotation or viewpoint focus functions.
(6) Rendering and editability: Check whether the lighting and color tone of rendered images are consistent, and whether project files can be opened in the corresponding 3D software and further modified.
4. Observations from One Test
Consider a miniature scene test with a detective agency theme. The results showed the following characteristics.
In terms of scene structure, the model largely reproduced the main components of the reference image: the side stairs, the second-floor office space, the first-floor coffee shop, and the character figures at the entrance. This indicates that the model had a good understanding of building levels and functional zoning.
In terms of key elements, the model omitted the most visible sign. This was the most important failure in the test. It suggests that when handling small but critical visual identifiers, the model may still ignore explicit requirements in the prompt.
In terms of character details, the character hairstyle differed significantly from the reference image, indicating that the fidelity of character details remains unstable.
For the interactive webpage, the model proactively added basic dragging, auto-rotation, scene annotations, and viewpoint focus, and retained a download entry and image switching. These features went beyond the literal requirements of the prompt, and whether to adopt them depends on project goals. For data visualization or sandbox-style projects, annotations and focus can provide practical value; for scenes that only need static display, they become an extra burden.
In terms of rendering, the color tone and filter treatment of the images were fairly consistent and could be used directly as presentation material.
5. A Practical Example: Structured Scoring Instead of Subjective Impressions
Subjective impressions are easily influenced by the first visual effect. It is better to convert each dimension into a score and record the reason for any deduction. The following example shows a simple weighted scoring method; the scores are for demonstration only:
RUBRIC = {"scene_completeness": 0.20,"key_elements": 0.30,"spatial_structure": 0.20,"detail_quality": 0.10,"interactivity": 0.10,"rendering_editability": 0.10,}def weighted_score(scores: dict) -> float:total = 0.0for name, weight in RUBRIC.items():value = scores.get(name, 0.0) # Each item is scored from 0 to 10total += weight * valuereturn round(total, 2)sample = {"scene_completeness": 9,"key_elements": 5,"spatial_structure": 8,"detail_quality": 6,"interactivity": 8,"rendering_editability": 8,}print(weighted_score(sample)) # Output: 7.1
In the example above, the missing key element receives a lower score, and its impact is amplified by the higher weight. This helps prevent extra features such as interactive webpages from masking a failure in the core task.
6. Quota and Cost Management
Most model services impose limits on calls or computing resources. When conducting systematic tests, plan the test matrix in advance. For example, use a fixed set of three to five prompts per theme and change only one variable per round. This controls consumption and also helps determine whether an improvement comes from a prompt adjustment. If quota is limited, spread tests across multiple days and record each result in the same table.
7. Adoption Recommendations
- Perform manual acceptance checks on generated content, especially for dimensions, structural safety, and signage compliance.
- For materials involving specific anime characters, film or television imagery, or brand logos, confirm copyright and scope of use first, and switch to original designs when necessary.
- For interactive webpages, first confirm whether target users need the features before deciding to keep or remove them.
- Build a fixed evaluation set containing simple, medium, and complex tasks, and retest periodically to observe capability changes.
The value of 3D content generation lies in shortening the path from an idea to a visual prototype. Only by making evaluation specific and reproducible can teams determine whether a model is truly suitable for entering the project workflow.
