Open benchmark · 1,280 objects
From reference image to 3D object.
Browse every attempt from one AI-assisted pipeline. Each entry puts the generated reference beside the resulting USDZ, including the misses.
CC BY 4.0 · Metric scale · Downloadable USDZ
The full benchmark
Compare every object.
Reference on the left, generated result on the right. Click either image to rotate the model, change its background, or copy a direct link.
Method
One pipeline, repeated 1,280 times.
Each object starts with fixed dimensions and a generated reference image. The pipeline builds it in Blender, packages it as USDZ, and runs technical checks on structure, scale, bounds, and placement. Nobody reviewed the results by eye, so judge the objects yourself.
Read the full method-
01
Specify
Set the size, category, and construction constraints.
-
02
Generate
Create the reference and build the object in Blender.
-
03
Package
Export a self-contained USDZ and run structural checks.
-
04
Compare
Keep the reference and result together, mistakes included.
Model bench
Compare the models, object by object.
GPT-6 Astra now covers all 1,280 objects, alongside the original builds and earlier Claude comparisons. Browse by model and category, compare each result with its shared reference, and open any build in 3D.
Explore the benchmarks- Pictures
- 1
- Models
- 2
- Ranked
- No
Use with judgment
The files are open. The results are imperfect.
The assets are available under CC BY 4.0. Credit “GenHome3D-1280 contributors” and link to the repository when redistributing or adapting them.
Automated checks cover file structure, scale, bounds, placement, textures, and checksums. They do not prove an object is correct, realistic, safe, or production-ready.