What it does

texturize takes a source image—grass, gravel, dirt—and generates new textures that match it. You can remix a single texture into variations, remake one texture in the style of another, mashup multiple sources, or enhance resolution using a higher-quality example. It’s built for 3D artists and graphics programmers who need more texture tiles without hand-painting them.

The interesting bit

The synthesis uses a hybrid approach: convolutional networks plus example-based optimization, processed in coarse-to-fine octaves. The library exposes both a command-line tool and a Python API that streams results octave-by-octave, which lets you bail out early or iterate on partial outputs.

Key highlights

• Four commands: remix, remake, mashup, enhance
• GPU (CUDA) or CPU, with float16/float32 precision control
• Configurable quality levels, output size, and random seed for reproducibility
• Colab notebooks for gravel and grass demos—no local install required
• AGPL license; installable via PyPI or Conda with provided setup scripts

Caveats

• PBR material support is listed as “soon” in the description, not yet implemented
• The README claims it’s “building as the highest-quality open source library available”—a self-assessment without external benchmarks cited
• Demo materials are BY-NC-SA licensed, which may complicate commercial use of examples

Verdict

Worth a look if you’re doing procedural content generation, game asset pipelines, or need to expand texture libraries without manual work. Skip if you need PBR workflows today or want a polished GUI; this is a research-grade tool with a CLI-first interface.