Evaluating AI-Generated 3D Scenes and Interactive Web Pages: Deliverables, Prompt Design, and Scoring Criteria

    This article explains how to evaluate AI-generated 3D scenes and interactive web pages. It covers why 3D generation is a useful test of model capability, how to separate three deliverable types, how to design bare and constrained prompts, and how to verify 3D projects across fidelity, structural integrity, detail, editability, and scale. It also outlines criteria for assessing interactive web pages, including basic interaction, information enhancement, and delivery support.

    Why 3D Generation Matters in Model Evaluation

    For some time, evaluations of large models have focused mainly on code completion, mathematical reasoning, and text understanding. As multimodal capabilities mature, 3D content generation is increasingly entering the evaluation scope. Compared with pure text output, 3D tasks simultaneously test spatial relationship understanding, structural decomposition, material and lighting design, and interaction logic implementation.

    A 3D scene requires a model to first convert an abstract description into geometry, and then organize that geometry into project files that can be edited, rendered, and interacted with. If any link in this chain fails, the final result will expose the problem. The value of 3D generation evaluation therefore lies in its ability to test both requirement understanding and execution.

    Splitting Deliverables into Three Categories

    A 3D generation task usually includes three types of deliverables, and each should be judged separately during evaluation.

    1. 3D model project: A source file that can be opened and modified in professional 3D modeling software. Focus on whether the object hierarchy is clear, whether parts can be edited individually, and whether materials are reasonable.
    2. Rendered images: Used to show the effect intuitively. Focus on composition, proportions, and recognizability of the main subject.
    3. Interactive web page: A 3D display page that can be opened in a browser. Focus on model loading, view control, and additional features.

    The quality of the three deliverables is not always synchronized. Rendered images may look attractive while the model project is missing parts; a web page may be feature-rich while the core model has insufficient fidelity. Looking at only one of these items can easily lead to partial conclusions.

    Prompt Design: From Bare Prompts to Constrained Prompts

    A bare prompt gives only the theme and general requirements without detailed restrictions. It reflects the model’s default understanding and is suitable for baseline testing. To compare the capabilities of different models, structured constrained prompts are needed. A reusable prompt usually includes the following elements:

    • Theme and scene: For example, a detective agency from a certain animated work, including the number of floors and functional zones.
    • Required element checklist: List the key objects that must appear, such as signs, furniture, and characters.
    • Style constraints: Such as voxel style, low-poly style, or realistic style.
    • Proportions and dimensions: Use consistent units and overall dimensions to facilitate horizontal comparison.
    • Delivery format: Specify file types, the number of rendered images, and web page requirements.

    Example:

    1. Theme: A miniature detective agency scene inside a three-story building
    2. Required elements:
    3. 1. An exterior wall sign with clearly readable text
    4. 2. An external side staircase connecting the first and second floors
    5. 3. The detective agency interior on the second floor, including an office desk and bookshelf
    6. 4. A coffee shop on the first floor, including a counter and seating
    7. 5. Two character models placed at the entrance
    8. Style: Voxel block style with a consistent block size
    9. Deliverables: An editable 3D project file + two rendered images + an interactive page that opens in a browser

    The required element checklist is the most critical part of the prompt. If a test omits such elements, it is impossible to determine whether the model truly understands the scene. Whether the model adds features that were not requested can reflect its inference ability, but this is a bonus item and cannot override the completion of basic requirements.

    Five Dimensions for Verifying 3D Models

    After receiving a model, it is recommended to score it item by item across the following five dimensions.

    1. Fidelity

    Fidelity focuses on how well the model matches the target scene. Missing key elements should lead to clear deductions. For example, if an iconic sign does not appear, even if other structures are complete, a full score should not be given. Recognizability often depends on a small number of key elements; without them, observers may find it difficult to recognize the original scene.

    2. Structural Integrity

    Check whether connections such as floors, stairs, and passages are reasonable. Common problems include floating stairs, rooms without entrances, and inconsistent floor heights. Structural integrity directly affects the usability of subsequent editing and animation production.

    3. Detail and Recognizability

    Whether the shapes of people or characters are recognizable is an important part of detail evaluation. Missing features such as hairstyle, clothing color, and proportions will weaken the overall effect. In voxel style, details are limited by block resolution, so evaluation should be based on style requirements rather than realistic standards.

    4. Project Editability

    After opening the model project, check whether objects are named by part, whether parts can be selected independently, and whether materials can be adjusted separately. Editability determines whether the model can enter subsequent workflows.

    5. Proportion and Scale

    Confirm whether the relative sizes of parts are reasonable. Scale errors are not easy to notice in static rendered images, but they become apparent when the model is imported into other scenes or used for animation.

    It is advisable to assign weights to each dimension and record the basis for judgments in a scoring sheet.

    Evaluation Points for Interactive Web Pages

    The features of an interactive web page often exceed the explicit requirements of the prompt. Evaluation can divide them into three categories: basic interaction, information enhancement, and delivery support.

    Basic interaction includes drag-to-rotate, zoom, and auto-rotation. Whether the page provides operation guidance when first opened is also part of the user experience.

    Information enhancement includes scene annotations and focus view lists. In smart visualization projects, annotations play a clear role. For example, in real-estate sandbox and urban modeling scenarios, building functions and area names can be labeled directly on the model to help users quickly understand the spatial layout.

    Delivery support includes model download and rendered image switching. These features give the page standalone delivery capability.