Evaluating 3D Generation Models: From Spatial Creation to Interactive Delivery

    This article reports a hands-on test of a 3D generation model asked to create a voxel-style detective agency scene and deliver editable project files, rendered images, and an interactive web preview. It examines structural accuracy, semantic completeness, interaction quality, and engineering usability, highlights failures in sign text and character fidelity, and proposes a five-dimension evaluation framework and a staged workflow for production use.

    In recent years, intelligent models have continued to improve in code generation, text question answering, and image creation. 3D content generation is now emerging as a new application direction. Compared with ordinary image generation, 3D creation requires more than making a model “look right.” The model must also understand spatial relationships between objects, structural hierarchy, material expression, viewing angles, and the final delivery format.

    In a recent test, I gave the model a relatively short prompt and asked it to create a voxel-style miniature scene themed around a classic detective agency. The requested deliverables included an editable 3D project file, rendered images, and an interactive version that could be viewed in a web browser. The final result was representative. The main building, floor relationships, and surrounding elements largely met expectations, and the model automatically added features such as model rotation, viewpoint switching, scene annotations, and resource downloads. However, there were still clear deviations in the key text signboard, character modeling, and some semantic details.

    This experience shows that evaluating a 3D generation model cannot rely only on whether the final screenshot looks “beautiful.” It also needs to be judged across multiple dimensions, including structural correctness, semantic completeness, interaction quality, and engineering usability.

    I. Test Objective: From Generating a Model to Delivering a Work

    The task was not simply to generate a 3D rendering. The goal was to obtain a relatively complete set of content assets, specifically:

    1. A voxel-style miniature scene with a clear theme;
    2. An editable 3D project file;
    3. At least two rendered images showing the overall effect;
    4. An interactive page that can be opened in a browser and rotated for viewing;
    5. As much restoration as possible of the key structures and iconic elements in the themed building.

    This task is much more complex than “generate a building model.” The model first needs to understand the theme in natural language, then break it down into different objects such as buildings, characters, stairs, signs, doors and windows, and storefronts. It must then establish 3D hierarchy relationships and finally organize the model into resources suitable for web interaction.

    From an engineering perspective, this is already close to a small content production workflow, rather than a traditional one-off question-and-answer interaction.

    II. Model Fidelity Is Acceptable, but Missing Key Semantics Still Affects Evaluation

    In the generated result, the restoration of the main building was basically in line with expectations. The scene included a multi-story building structure, with stairs on the side leading to the upper floor. The ground floor contained a coffee shop, while the upper floor was used as a detective agency. Two character models with a narrative connection were also placed at the entrance.

    These elements show that the model has a certain ability to decompose a scene. It did not simply interpret the task as “make an ordinary house.” Instead, it identified the main components of the themed building and placed different functional areas in relatively reasonable positions.

    However, the model omitted the most important text signboard on the building wall. This type of problem is very common in 3D generation: the model can understand that “this is an agency here,” but it does not necessarily generate the specified text accurately. There are three main reasons:

    1. Text generation itself has high error rates

    Whether in 2D images or 3D models, text is harder to generate than ordinary textures. The model must simultaneously handle character shapes, arrangement direction, font style, material fitting, and spatial perspective. If any one of these links deviates, the signboard may become garbled, lose characters, or be missing entirely.

    2. Models tend to prioritize overall structure

    When a prompt includes requirements for buildings, characters, style, floors, rendering, and delivery formats at the same time, the model often prioritizes volume relationships and major objects, while placing details such as text and decorative patterns at a lower priority.

    Semantically, the model has already generated a detective agency, so it can be considered to have understood the theme. But if the signboard that immediately confirms the identity is missing, the recognizability of the work still decreases.

    Therefore, when accepting a 3D generation result, we should not only ask “does it look like it?” We should also establish a checklist of key elements. For example:

    • Whether the main building exists;
    • Whether the floor relationships are correct;
    • Whether the entrance and stairs are reasonable;
    • Whether the main characters appear;
    • Whether the iconic signboard exists;
    • Whether the signboard text is readable;
    • Whether each functional area matches the prompt.

    As long as one of these items is a high-priority element, it should not be overshadowed by the completion of ordinary decorative details.

    III. Shorter Prompts Better Reveal the Model’s Raw Capabilities

    This test deliberately did not attach additional skills or provide complex reference materials. It directly used a bare prompt. The advantage of doing so is that it allows observation of the model’s default capabilities when human assistance is lacking.

    The prompt roughly contained the following information:

    • Create a voxel-style 3D miniature scene;
    • The theme is a detective agency;
    • Reference the architectural relationships in classic detective stories;
    • Highlight the agency as the visual center;
    • Deliver editable project files and rendered images.

    This wording was enough to explain the creative direction, but not enough to cover all acceptance requirements. To improve the completion rate of key details, it is recommended to split the prompt into five parts: task objectives, structural requirements, visual requirements, delivery requirements, and acceptance criteria.

    For example:

    1. Please create a voxel-style 3D miniature building scene.
    2. Theme:
    3. - A multi-story building containing a ground-floor coffee shop and an upper-floor detective agency;
    4. - The agency is located at the visual center;
    5. - External stairs on the side of the building lead to the second floor.
    6. Must include:
    7. - A Chinese-language agency signboard on the building exterior wall;
    8. - Ground-floor store entrance, windows, and awning;
    9. - Two character models at the entrance;
    10. - Recognizable floor zoning;
    11. - A base and environmental decorations suitable for a miniature scene.
    12. Visual requirements:
    13. - Build with regular voxel blocks;
    14. - Use soft lighting and moderate depth of field;
    15. - Keep the building structure clear and avoid an overly complex background.
    16. Delivery requirements:
    17. - An editable 3D project file;
    18. - At least two rendered images from different viewpoints;
    19. - A web preview version that supports rotation, zoom, viewpoint reset, and object annotations.
    20. Acceptance criteria:
    21. - The signboard text must be complete and readable;
    22. - Functional areas of the building must not intersect with each other;
    23. - The proportions of characters and building should be basically coordinated;
    24. - All delivered files must open normally.

    The point of this type of prompt is not to make it longer, but to let the model know which requirements are “must complete” and which are only “style preferences.” If priorities are not distinguished, the model tends to process all requirements evenly, causing key elements to be omitted.

    IV. Automatically Generated Web Interaction Reflects Product Awareness

    Compared with the 3D model itself, the automatically generated web interaction version was more surprising. The original requirement was only that the file could be opened in a web page, but the final page not only supported basic dragging, rotation, and zooming, but also added several display-oriented features.

    1. Model Rotation and Zoom

    This is the basic capability of a 3D web preview. Users can view different sides of the building through mouse or touch operations, and can also zoom in to observe local details such as the signboard, characters, and stairs.

    For a miniature scene, rotation is especially important. A single rendered image usually shows only one angle, while an interactive model can help users verify whether the building truly has a complete structure and avoid the problem that “the front looks good, but the back is not modeled.”

    2. Scene Annotations

    The web page added annotations for the main components of the model, such as the main building, coffee shop, and characters. Although annotation may seem simple, it is very useful in smart displays, city sand tables, campus planning, architectural presentations, and educational demonstrations.

    An effective annotation system usually needs to include:

    • The binding relationship between annotation points and 3D objects;
    • Label readability under different viewpoints;
    • Display strategies under occlusion;
    • Focus effects after clicking a label;
    • Annotation toggles and batch hiding capabilities.

    If text is simply fixed on the page, labels may drift after the model rotates. A more ideal approach is to bind annotations to 3D coordinates and dynamically calculate screen positions based on the camera position.

    3. Viewpoint Focus

    The sidebar lists the main objects in the scene. When the user clicks an object, the camera automatically moves to a suitable position. This feature can significantly lower the barrier to using 3D models, especially for non-professional users.

    From an implementation perspective, viewpoint focus usually requires recording for each object:

    • Bounding box;
    • Center point;
    • Recommended viewing distance;
    • Recommended camera direction;
    • Focus animation duration.

    If object sizes vary greatly, a uniform camera distance will make small objects hard to see and large objects impossible to display completely. Therefore, the camera position needs to be dynamically adjusted according to the object bounding box.

    4. Resource Download and Image Switching

    The web page also provides model file downloads and switching between multiple rendered images, showing that the delivery approach has extended from “displaying the model” to “managing assets.” For real projects, the preview page is not only a display interface but can also serve as a lightweight content delivery entry point.

    However, for actual production environments, file versions, resource descriptions, update times, format information, and licensing notes should also be added, so that users can determine the purpose of files after downloading.

    V. Character Modeling Issues Show That Local Semantics Still Require Manual Review

    Although the number and positions of characters in the scene were basically reasonable, there were certain deviations in hairstyles, clothing, and proportions. Character models account for a small proportion of the overall scene, but they easily become visual focal points, so local errors are quickly noticed by users.

    This shows that current 3D generation models are usually better at “object category recognition” than at “character identity restoration.” The model can generate a person standing at the entrance, but it may not reliably restore the hairstyle, clothing, and pose of a specific character.

    For this type of problem, a staged generation strategy can be adopted:

    1. First generate the building and environment;
    2. Then generate the main characters separately;
    3. Review the characters’ proportions, poses, and clothing;
    4. Finally merge the characters into the scene as independent objects.

    Compared with asking the model to complete all content at once, staged processing makes it easier to locate problems and facilitates subsequent modifications. Characters should be kept with independent names and independent object structures as much as possible, so that they can be separately hidden, replaced, and adjusted in a 3D editor or web page.

    VI. Rendering Quality Is Good, but Aesthetics Cannot Replace Structural Accuracy

    The final rendered images used soft lighting and a somewhat illustrative visual style, and the overall look was fairly consistent. Appropriate filters can enhance the fun of voxel models and can also weaken some shortcomings in geometric details.

    However, in technical acceptance, “visual effect” and “model quality” need to be evaluated separately:

    • Whether the lighting is natural;
    • Whether shadows can reflect spatial relationships;
    • Whether materials match the voxel style;
    • Whether the camera angle highlights the main subject;
    • Whether the background interferes with the building;
    • Whether objects intersect with each other;
    • Whether the model has a truly editable structure.

    A good-looking image only proves that the rendering result is usable. It does not prove that the project file structure is reasonable. If the model consists only of a large number of non-editable merged meshes, subsequent modification of the signboard, replacement of characters, or adjustment of stairs will become difficult.

    Therefore, project file acceptance should at least check:

    • Whether objects are grouped by function;
    • Whether buildings, characters, and decorations can be edited independently;
    • Whether materials are reasonably named;
    • Whether cameras and lights are preserved;
    • Whether there are obvious duplicate, floating, or intersecting objects;
    • Whether the file displays correctly after being reopened.

    VII. Building a More Complete Evaluation System for 3D Generation

    Through this test, the evaluation of intelligent 3D creation can be divided into five dimensions.

    1. Semantic Understanding

    Whether the model understands the theme and can identify the main buildings, characters, and functional areas.

    2. Spatial Reasoning

    Whether the front-back, up-down, inside-outside, and connection relationships between objects are reasonable; whether stairs lead to the correct floor; whether doors and windows are in suitable positions.

    3. Detail Restoration

    Whether signs, text, character modeling, decorations, and materials meet requirements.

    4. Engineering Delivery

    Whether editable files can be delivered, whether the scene structure is clear, whether resources are complete, and whether the files are convenient for secondary modification.

    5. Interactive Product Capability

    Whether it supports rotation, zoom, focus, annotation, viewpoint switching, resource download, and first-use guidance.

    Among these five dimensions, the first two determine “whether the model is correct,” the third determines “whether the work looks right,” the fourth determines “whether it can continue to be used,” and the fifth determines “whether users are willing to use it.”

    VIII. Conclusion: 3D Generation Is Moving from Single-Point Creation to Complete Delivery

    Overall, the biggest progress shown in this test is not only that the model can generate a voxel building, but that it began to proactively fill in peripheral capabilities such as web interaction, scene annotation, and resource organization.

    This means that the output form of intelligent models is changing. In the past, they mainly delivered text, code, or images. In the future, they are more likely to directly deliver a complete small project containing a model, rendered images, an interactive page, and explanatory information.

    Of course, current results still cannot fully replace manual acceptance. Key text, character details, proportional relationships, and engineering structures all need review. Especially in commercial displays, educational content, and formal projects, the first automatically generated version should not be published directly.

    A more reasonable way of working is to treat the model as an “efficient first-draft producer” and “interactive prototype builder,” then complete the final delivery through clear prompts, structured acceptance checklists, and manual revision. For developers, the real question is not whether the model can generate a perfect work in one pass, but whether it can significantly shorten the entire process from concept and modeling to interactive presentation.