Evaluating LLM-based 3D Voxel Modeling: From Prompt to Web Acceptance
This article evaluates an LLM's ability to generate 3D voxel models and interactive web pages using natural language. It details the testing configuration, prompt engineering, acceptance criteria for spatial structures and visual elements, and an HTML interaction checklist. The review highlights both successful outcomes and current limitations, providing developers with a practical framework for assessing and scaling AI-driven 3D content generation.
1. Why Record Configurations and Acceptance Criteria
The effectiveness of LLM-generated 3D content heavily relies on the environment, prompts, and criteria. Vague statements like “good results” or “obvious flaws” do not help readers determine if the conclusions apply to their own projects. Therefore, this test is divided into three parts: inputs (configurations and prompts), outputs (modeling results and web delivery), and evaluation (acceptance checklists and failure cases).
The core objective is to test the model’s ability to create 3D content using natural language without preset skills or plugins. It must simultaneously handle three tasks: ensuring spatial structure is reasonable, specific visual elements are preserved, and deliverables are directly usable.
2. Test Configuration and Prompts
To reproduce the test, the following information should be recorded and supplemented if missing from the original logs:
- Model name and version
- Calling method (web, API, or client) and whether web search or tool calling is enabled
- Blender version and operating system
- Plugin usage (no plugins were used in this test)
- Generation parameters (temperature, max output length, etc.), with defaults noted for unadjusted items
- Daily API quotas and whether tasks were split across multiple sessions
The core prompt used in this test was:
Create a 3D voxel miniature landscape themed “Mori Detective Agency” using Blender, and open the file in HTML.
For reusability, the prompt can be templatized into variables:
Create a 3D voxel miniature landscape themed “{theme}” using Blender. It must include {structural elements} and preserve {visual logos}. The number of characters is {character count}, and their positioning is {positioning description}. Output the Blender project file and renderings, and open the file in HTML.
Only one variable should be changed at a time to isolate the source of any changes in the output.
3. Modeling Results and Acceptance Methods
During acceptance, inspect item by item rather than just evaluating the overall appearance.
- Spatial Structure: Check if the side stairs connect the first and second-floor main entrances, if the street-level shop and second-floor partitions are clear, and if the height ratios of each floor are reasonable. The validation method involves switching to the orthographic view in Blender to verify the number of stair steps and floor heights against the design.
- Signature Elements: Check if the vertical signboard exists and if the text and position are correct. The validation method is to inspect both the rendering and the project file. Elements that appear only in the rendering but are missing from the project file should be marked as a failure.
- Character Details: Check if the character count, positioning, hairstyles, and clothing are recognizable. The validation method is to zoom in to the character scale and record which features were simplified.
Conclusion: The spatial structure is generally reasonable and aligns with common architectural layouts. The vertical signboard is missing, and character handling tends to be abstract. Simplification of characters in a voxel style is acceptable, but commercial-grade assets would still require manual refinement.
4. HTML Interaction Delivery Acceptance
The prompt only requested opening the file in HTML; most interactive functions were proactive extensions by the model. While these extensions have practical value, each must be validated individually. A complete-looking page does not guarantee that every function works.
| Feature | Test Performance | Acceptance Method |
|---|---|---|
| Drag rotation & auto-easing | Generated | Release mouse after dragging; check if model continues to rotate slowly |
| Element annotation | Generated, distinguishes characters/buildings | Click annotations to verify pointing accuracy |
| View focusing | Generated, switches on viewport name click | Click names to confirm camera alignment |
| Model download | Generated | Reopen in Blender to check for missing materials/geometry |
| Multi-image switch | Generated | Switch images to confirm match with current view |
| First-load guide | Generated | Clear cache, open first time, confirm guide appears |
These functions were observed during testing, but no screenshots or screen recordings were saved in the original logs. Evidence needs to be supplemented during re-testing.
5. Failure Cases and Limitations
- Missing signboard: The model generated the main building and characters but failed to embed the signature signboard. The output indicates it organized the structure based on the overall architectural style without prioritizing specific cultural symbols. The exact reason is unverified. Re-testing could isolate the signboard as a separate requirement at the beginning of the prompt to observe improvements.
- Simplified character details: Hairstyles and clothing were compressed into a few color blocks. Characters are distinguishable but lack sufficient detail.
- Large scenes untested: Due to daily API quota limits, stress testing entire building complexes and city blocks was postponed.
6. Copyright and Compliance Notes
“Mori Detective Agency” and related characters and scenes originate from a well-known anime. Personal learning and internal testing face different licensing requirements than public release or commercial use. If you intend to use the generated models, renderings, or web pages for promotion, sales, or public display, you should first confirm copyright ownership and authorization scope. In commercial scenarios like real estate sales or digital twins, avoid directly using copyrighted character images.
7. Conclusion
In this bare-prompt test, the model successfully converted a natural language prompt into a structurally reasonable voxel building accompanied by an interactive web page. Its shortcomings lie in the preservation of signature elements and character details. Developers looking to use such workflows in formal projects should establish fixed prompt templates and acceptance checklists, evaluate item by item after each generation, and record failure cases. Only after multiple re-tests yield stable results is it worth expanding the scope of use.
