Blender benchmark

The same house brief, different models.

The benchmark compares how AI models create a complete furnished bungalow in Blender, from layout and editable geometry to materials, renders, and preparation for an interactive tour.

Results
2
Demos to try
2
Method
Shared Blender prompt

In brief

  • Every model receives the same assignment to create a complete furnished bungalow in Blender.
  • The brief includes layout, editability, visuals, and a 3D export; the current ranking uses renders only.
  • The comparison includes GPT-6 Astra and Fable 5.1; the refined Astra High appears separately below.
  • The displayed ranking is based on final render quality, without a new review of layout or editability.

All results

Results and 3D viewer

Two Blender outputs ranked by their final renders. Each can be explored in 3D and downloaded as the original .blend file.

Bungalow benchmark results Ranked by the visual quality of the final renders
Preview Model Type Parameter count / requirements Result Limitations Note 3D model
1GPT-6 Astra Closed — Clean, coherent visuals Working 3D preview Bright, sometimes flat exterior - Viewer Download .blend 6.2 MB
2Fable 5.1 Closed — Detailed interior Working 3D preview Darker renders; simplified greenery - Viewer Download .blend 20.2 MB

Every model supports orbiting, zooming, and walking through the house.

Sources and run data

dates, prompt, run type, checkpoints, revisions, and memory data

Only documented data is included. A run date, prompt version, and run type are shown only when they were recorded; historical gaps are not estimated. Q4 weight size is not total RAM usage; cache or state at 100K is a separate figure for one sequence and the stated runtime.

GPT-6 Astra
Generated
Prompt
bungalow-master-en-v1
Run type
multi-prompt
Fable 5.1
Generated
Prompt
bungalow-master-en-v1
Run type
multi-prompt

Refined version

Result refined with follow-up prompts

A separate result created and refined through multiple prompts, outside the main ranking.

  • Multi-prompt · High effort

GPT-6 Astra High

Casa Oliva — a refined furnished bungalow with materials, a garden, and an interactive walk. The original Blender file is available to download.

Method

What the benchmark tests

Every model starts with the same prompt. Ultra effort is used by default unless a result explicitly states otherwise. This ranking comes from a visual comparison of final renders; layout and editability of the new .blend files were not reviewed here. Follow-up fixes must remain within the original brief and the run is labeled multi-prompt.

Original assignment prompt

The model received the English prompt in the original wording shown below.

Create a complete, fully furnished, single-story three-bedroom bungalow in Blender, with a small garden, terrace, and pergola, combining modern design with Italian elegance. Aim for photorealistic materials, lighting, and details, with a practical layout, independently accessible bedrooms, and correctly positioned furniture. Model the entire house, including full walls, ceilings, and roof. Keep door leaves separate from their frames, with correctly positioned hinge pivots or defined sliding directions, ready for opening and closing in a WebGL viewer. Critically inspect the model and fix mistakes throughout the process. Deliver an editable .blend file and interior and exterior renders.

FAQ

Bungalow benchmark FAQ

What does the Bungalow benchmark test?

It tests an AI model's ability to design and build a complete furnished bungalow in Blender from a shared brief.

How is the result evaluated?

The review covers the house layout, editable geometry, materials, render quality, and operation of the exported model in the 3D viewer.

Can the models be inspected interactively?

Yes. All three results open in the shared 3D viewer with orbit and walk modes.

Are the original Blender files available?

Yes. Below each Viewer link you can download the original, unchanged .blend file. The viewer uses a separate web export.

How are the results ranked?

By the visual quality of the final renders. This ranking does not assess the layout or editability of the new .blend files.