How We Evaluate AI 3D Generators
Last reviewed 2026-11-08
PromptToProp currently evaluates AI 3D generators from published sources: vendor pricing pages, documentation, license terms, model cards and clearly attributed third-party reports. We have not yet run our own generator tests. This page explains how we work today and exactly how testing will work when it starts, so you can hold us to it.
How pages are evaluated today
Every page is labeled by the kind of evidence behind it.
| Label on the page | What it means |
|---|---|
| Based on published sources | Written from documentation, pricing, terms and attributed reports. No original tests from us. |
| Measurements not run yet | The page explains what to compare and how, but holds no results from us. |
| Original test data | Results from a logged test, with date, version, plan and method. No page carries this label yet. |
When we recommend a generator on a source-based page, the reasoning rests on things that can be checked: what a plan costs, which file formats it exports, what the license permits, and what its documentation says it does. We attribute any quality claim to whoever made it, and we point out when that party sells the product.
Prices, credit allowances and license terms change often. Each is dated, and every page shows when it was last reviewed.
What we weigh
Fit for the job. A generator that suits one-off print figurines may be a poor choice for a game prop library. We recommend by job, not with a single overall winner.
Total cost. Credits burned on retries and hours spent on cleanup can outweigh the subscription. Our credit cost calculator works that through with your own numbers.
Rights. Free tiers often come with different terms from paid plans. We read the terms and summarize them in the license matrix and our licensing guides.
Output you can use. Export formats, documented mesh and texture options, and published limits.
Downsides. Every review states documented weaknesses. If we cannot find any information on something, we say that.
The planned test protocol
The site’s central project is same-prompt testing across tools, published as it is completed in the test hub. None of the following has been carried out yet.
Same inputs. A fixed prompt set and scoring rubric covering props, characters, stylized assets, creatures and image input, published before results so the choice of prompts cannot be tuned afterward.
Comparable conditions. Named paid plans, all runs within one short date window, tool and model version recorded for each run.
Logged fields for every run.
- Prompt or input image, tool, plan, model version, date
- Credits used, generation and queue time, failed runs and retries
- Triangle count, UV island count, watertightness and defects, from scripted mesh analysis rather than judgment
- Texture channels supplied, and visible baked lighting or seams
- Cleanup minutes to reach a defined usable state
Real destinations. Exports brought into Godot, Unity and Unreal, and print candidates run through slicers with pass or fail criteria stated in advance. A physical print is only claimed where one was made.
Raw evidence kept. Output files archived where each tool’s terms allow redistribution, with screenshots and a dated results table.
Retests. Major model releases trigger a retest of the affected pages. Changes are logged in the corrections and retest log.
What will not happen
No tool pays for placement or a better result. Vendors will not see results before publication or be able to approve them. Commission rates play no part in rankings, and we explain how we handle commercial relationships in our affiliate disclosure.
We also will not call something tested when it was not. If a page says it rests on published sources, that is all it rests on. Our editorial policy covers sourcing and how AI tools are used in producing the site.