ML-powered logo maker for easily creating custom logo designs

The key challenges

‍

‍

1. Implementing good colour matching of particular elements within generated logos.
2. Allowing users to fine-tune selected logos.
3. Providing well-looking generated logos in vector graphic format.

‍

The project boiled down to creating a proof of concept version (POC) of the Python application in order to demonstrate its feasibility and verify a practical potential of the concept.

‍

In brief, the main user flow looks as follows: user inputs name of the business, selects industry, chooses a few pre-designed styles from templates, picks up colour schemes. These several steps were supposed to lead to the automated generation of a logo with a watermark.A fundamental job that determined the success of the project was designing the model of logo generator's engine.

‍

This came down to the following flow:

‍

1. At the beginning of the process, a batch of logos with random parameters was created.
2. From this batch, those which broke some base design principle were filtered out i.e. those without sufficient contrast between colours.
3. Finally, a Machine Learning classifier, trained on examples hand-labeled as good or bad, filtered out those logos that didn’t reach a certain quality.

The model used as machine learning classifier was a neural network created in TensorFlow. This framework exposes to programmer low-level API that allows having better control over the architecture of the neural network.