The idea of fully delegating 3D graphics creation to language models sounds enticing. While some developers are setting up direct interaction via the MCP (Model Context Protocol), others choose a more straightforward path: generating ready-made Python scripts for Blender and executing them directly within the software.
How the Method Works
The premise of the experiment is simple: a text prompt describing a desired medieval fantasy object in a retro low-poly style is sent to the neural network. The output is ready-to-use code that, when executed inside Blender, builds a 3D mesh and assigns basic materials.
Generating the script usually takes a couple of minutes, while the generation in the editor takes mere seconds. However, the practical value of the result directly depends on the complexity of the geometry.
What Works Well and Where the AI Struggles
Testing various game assets revealed clear patterns:
- Simple geometry and furniture: Tables, mana crystals, basic round shields, and simple towers turn out quite usable as greybox placeholders or in stylized indie projects.
- Complex mechanics and organic shapes: Attempts to create detailed weapons, an ornate cannon, a griffin, or an anatomically correct horse lead to severe distortion. The mesh stretches, proportions break, and elements overlap incorrectly.
- High-poly and fine detailing: The model is not yet capable of generating a complex high-poly model with procedural metal and grime textures via a simple script.
Language models are great at assembling basic primitives, but they have a poor understanding of physical surface interactions, complex cloth deformation, and anatomical principles.
Conclusion: Is It Worth Using in Development?
The "prompt → Python code → Blender" pipeline is currently suitable exclusively for rapid prototyping and creating rough low-poly base meshes that will inevitably need manual refinement.
To produce higher-quality static 3D models for production today, it is more effective to combine 2D concept generators (like Krea or Nano Banana Pro) with subsequent 3D conversion using specialized AI services (such as Meshy).