From 3D Generation to Interactive Delivery: How to Evaluate Next-Generation Multimodal Models

    This article examines a test of a newly released general-purpose multimodal model. The model received a plain-language request to build a voxel-style 3D scene of a detective agency. The article covers what the model got right, what it missed, why a bare prompt can still produce a complex scene, how to write structured prompts, and how to evaluate 3D outputs for visual quality, structure, semantic accuracy, spatial reasoning, interactivity, and delivery reliability.

    Developers are judging multimodal models less by whether they can write code and more by whether they can complete a full digital work. Among the scenarios that measure a model’s overall ability, 3D modeling, spatial relationship understanding, editable asset generation, and packaging the output as an interactive web page have become especially important.

    Recently, I tested a newly released general-purpose multimodal model. I did not attach extra skills or provide complex reference files. I entered one direct natural-language request: create a voxel-style 3D miniature scene themed on a classic detective agency, and deliver editable project files and rendered images.

    The result was somewhat surprising, and it also exposed several issues worth analyzing. The model largely reproduced the spatial structure of the target building. It generated stairs, multiple floors, a ground-floor shop, and character models. It also output rendered images and editable 3D project files. More unexpectedly, the model automatically generated an interactive web page for browsing the work. The page included model rotation, zooming, camera focus on specific objects, scene annotations, image switching, and file downloads.

    This test suggests that the value of current 3D generation models no longer lies only in producing a model. They are beginning to deliver a relatively complete content experience.

    1. Does the Result Meet the Core Requirements?

    In this test, the building’s outline and main components were fairly complete. The multi-story structure, exterior staircase, ground-floor shop, and the person in front of the entrance all appeared in the final work.

    From a spatial reasoning perspective, the model completed at least five tasks:

    1. It identified the building’s theme and functional zones from the text.
    2. It understood the vertical relationship between floors.
    3. It organized people, the building, and the street environment into one scene.
    4. It used a consistent voxel visual language.
    5. It output 3D project assets that can still be edited.

    The result cannot be judged simply as “similar” or “not similar.” The model omitted the most recognizable sign text in the scene. This is a typical loss of a key semantic element.

    3D generation often reproduces the overall outline while skipping key elements that the user did not explicitly repeat. For a person, the building sign may be the core of recognizing the scene. For the model, it may be just one of many local details. The evaluation should therefore distinguish between:

    • Structural accuracy: whether the building levels, proportions, and layout are reasonable;
    • Semantic completeness: whether core identifiers and thematic objects are all present;
    • Visual consistency: whether materials, lighting, color, and style are unified;
    • Editability: whether the model consists of logical objects and hierarchy;
    • Delivery completeness: whether the project files, renders, and preview page are usable.

    If you only look at the rendered images, the work appears finished. From a delivery standpoint, however, the missing sign means the requirements were not fully met.

    2. Why a Bare Prompt Can Still Produce a Complex Scene

    This test did not provide modeling steps, an object list, or instructions for the web page. It used a short natural-language prompt. The model still produced a fairly complete result, mainly because it can break down tasks.

    A complete 3D creation instruction can usually be split into four layers:

    1. Theme layer

    Define what the work represents, such as a detective agency, a street of shops, an exhibition space, or an urban model.

    2. Structure layer

    Define which subjects appear in the space and how they are arranged. For example, how many floors the building has, where the entrance is, and where the people are located.

    3. Style layer

    Decide whether to use realistic, cartoon, low-poly, voxel, or industrial design styles.

    4. Delivery layer

    Specify the final files required, such as an editable project, rendered images, a web preview, or a printable model.

    When a prompt covers all four layers, the model is more likely to treat the task as a complete project rather than a single image.

    A bare prompt works well for testing a model’s basic capability, but it is not ready for production. It usually lacks clear acceptance criteria, and it does not specify object naming, model hierarchy, material organization, or file structure. For experiments, this approach is fast enough. For team collaboration or commercial projects, the prompt needs more structure.

    3. Why the Interactive Web Page Matters

    The most valuable part of this result may not be the 3D model itself. It may be the interactive web page that the model generated automatically.

    Beyond displaying the model, the page included several common features:

    • Drag to rotate the model;
    • Scroll to zoom the view;
    • Automatic rotation for presentation;
    • Buttons that focus the camera on specific objects;
    • Descriptions of scene elements;
    • Switching between different rendered images;
    • Downloads for the project file and image assets;
    • An operation hint shown on first open.

    The model added many of these features on its own. They were not all specified in the original request. This suggests the model is beginning to understand the difference between delivering a work and making that work usable.

    A 3D file is only an asset. An interactive page provides an entry point. For users who are unfamiliar with professional modeling software, a web preview lowers the barrier to access. For project reports, design reviews, client presentations, and digital exhibitions, this kind of packaging is also closer to real workflows.

    In scenarios such as real estate showcases, campus planning, urban models, and industrial equipment explanations, scene annotations and camera focus are especially useful. Users do not have to search for objects in a complex 3D editor. They can click labels such as “Entrance,” “Main Building,” or “Parking Area” and go directly to the target area.

    4. Automatically Added Features Are Not Fully Reliable

    The ability to add features automatically is useful, but it can also create new problems.

    First, automatically generated object names may be inaccurate. In this test, the page labeled one character with a specific role, but the character’s design did not clearly reflect that identity. Its hairstyle and facial features were noticeably off. This shows that the model can build a relationship between an object and its label, but it cannot guarantee that the visual result matches the label.

    Second, the web features may be superficial. A button that moves the camera does not mean the scene structure is well organized. If the model objects are not clearly named, later edits and maintenance will still be difficult.

    Third, an automatically generated interactive page needs compatibility and security checks. At a minimum, verify that:

    • The page opens correctly in common browsers;
    • Model file paths are relative;
    • The page shows a message when resources fail to load;
    • The page works on mobile devices;
    • Large models do not cause the page to lag;
    • Download links point to the correct files;
    • The page does not include unnecessary external requests.

    A web page that looks functional is only the first step. Before it enters a project, it still needs formal acceptance testing.

    5. How to Improve 3D Generation Prompts

    To make results more stable, expand a one-sentence request into a structured requirement. You can organize the prompt around the following parts:

    Scene goal

    Specify the theme, audience, and intended use. For example, a web display, classroom teaching, a project presentation, or 3D printing.

    Required objects

    List the core elements that must not be missing, such as the building sign, entrance, stairs, people, light boxes, and street base.

    Spatial relationships

    Describe how the objects relate to each other. For example, the building sits in the center of the base, the stairs are on the side, people stand near the entrance, and the shop is on the ground floor.

    Style and constraints

    Specify voxel size, number of colors, lighting, camera angles, and whether complex textures should be avoided.

    Delivery checklist

    Specify the project file, rendered images, web preview, and asset folder. Also ask the model to name files logically.

    Acceptance criteria

    Ask the model to check its own output after generation. It should confirm that core objects exist, that sign text is visible, that the model opens, and that the web features work.

    For example, the requirement could be written as follows:

    1. Create a voxel-style, multi-floor miniature scene of a detective agency.
    2. Required elements:
    3. 1. A building sign with clear, readable text;
    4. 2. An exterior staircase leading to the second floor;
    5. 3. A shop on the ground floor and office space on the second floor;
    6. 4. Two character models;
    7. 5. A building base and basic street environment.
    8. Deliverables:
    9. - An editable 3D project file;
    10. - Two rendered images from different viewpoints;
    11. - An interactive page that can be opened in a browser.
    12. The web page must support:
    13. - Rotation and zoom;
    14. - Automatic rotation;
    15. - Labels for key objects;
    16. - Camera focus on an object when it is clicked;
    17. - Switching between images;
    18. - File downloads.
    19. After completing the task, check the result against three items: core objects, file completeness, and web page functionality.

    Compared with a simple description, this format reduces omissions and makes the output easier to accept.

    6. Metrics for Evaluating 3D Generation Models

    To judge whether a model has practical 3D creation ability, use several metrics rather than visual impression alone.

    Visual quality

    This covers composition, proportion, materials, lighting, and overall style. The quality of the renders determines the first impression.

    Structural quality

    Check whether the model has a sensible object hierarchy, naming, and transform information. A model that looks good but cannot be edited has limited engineering value.

    Semantic accuracy

    Verify that the model includes the objects the request specified, especially titles, identifiers, and thematic symbols.

    Spatial reasoning

    Observe whether floor relationships, occlusion, foreground and background, and the scale of people relative to buildings are reasonable.

    Interaction quality

    Check whether the web page helps users understand the scene, rather than only placing the model inside a page.

    Delivery reliability

    Confirm that the files are complete, paths are correct, resources are not missing, and the result can be reproduced in different environments.

    7. From Demonstration Capability to Production Capability

    3D generation models can now handle proof-of-concept work, visual sketches, and lightweight display pages reasonably well. Production use still requires solving several practical problems.

    The first is text generation. Signs, labels, and interface text still often have missing content, typos, or poor legibility. Important text usually needs separate handling in post-production.

    The second is geometric precision. Generation models are good at conveying the overall look, but they are not always suitable for engineering models that require accurate dimensions, joints, and assembly tolerances.

    The third is asset standards. Production projects usually require consistent naming, materials, hierarchy, units, and version control. These cannot be fully left to automatic generation.

    The fourth is performance. Web-based 3D display needs control over polygon count, texture size, and file size. Otherwise, the page may load slowly or lag on low-performance devices.

    Overall, this test does not show whether a single model deserves a perfect score. It shows that 3D generation is moving from producing single pieces of content toward a combined workflow that covers model assets, visual rendering, and interactive delivery. The model’s ability to add features on its own suggests that its understanding of tasks is improving. Missing key elements and mismatches between character labels and designs, however, show that a single generation cannot be fully relied on.

    A more reliable approach is to first generate a draft with natural language, then add the core objects with a structured prompt. Next, check the project files, correct the text and model hierarchy, and finally test the web interaction, resource loading, and cross-device compatibility. Only after this process can 3D generation results move from demonstration to real application.