[Paper Review] LoGANv2: Conditional Style-Based Logo Generation with Generative Adversarial Networks

Cedric Oeldorf, Gerasimos Spanakis|arXiv (Cornell University)|Sep 22, 2019

Generative Adversarial Networks and Image Synthesis22 references34 citations

TL;DR

This paper extends StyleGAN with conditional latent spaces and a conditional WGAN-GP loss to enable higher-resolution, controllable logo generation, using object-classification and ResNet-based labels for conditioning. It evaluates unconditional vs conditional models on logo data and analyzes trade-offs between quality, diversity, and adherence to learned classes.

ABSTRACT

Domains such as logo synthesis, in which the data has a high degree of multi-modality, still pose a challenge for generative adversarial networks (GANs). Recent research shows that progressive training (ProGAN) and mapping network extensions (StyleGAN) enable both increased training stability for higher dimensional problems and better feature separation within the embedded latent space. However, these architectures leave limited control over shaping the output of the network, which is an undesirable trait in the case of logo synthesis. This paper explores a conditional extension to the StyleGAN architecture with the aim of firstly, improving on the low resolution results of previous research and, secondly, increasing the controllability of the output through the use of synthetic class-conditions. Furthermore, methods of extracting such class conditions are explored with a focus on the human interpretability, where the challenge lies in the fact that, by nature, visual logo characteristics are hard to define. The introduced conditional style-based generator architecture is trained on the extracted class-conditions in two experiments and studied relative to the performance of an unconditional model. Results show that, whilst the unconditional model more closely matches the training distribution, high quality conditions enabled the embedding of finer details onto the latent space, leading to more diverse output.

Motivation & Objective

• Address the challenge of high-modality logo data in GANs and improve training stability at higher resolutions.
• Introduce a conditional extension to StyleGAN to gain controllability over generated logos.
• Develop and compare two conditioning strategies derived from automated label extraction.
• Assess how conditioning affects quality, diversity, and learning of multi-modal logo distributions.

Proposed method

• Adopt progressive training and a style-based generator inspired by StyleGAN to stabilize high-resolution logo synthesis.
• Incorporate conditioning by embedding class-conditions into the intermediate latent space prior to the mapping network.
• Update the GAN loss to a conditional WGAN-GP to integrate conditioning labels into discriminator learning.
• Extract visual-class conditions using two approaches: object-classification-based labeling and ResNet feature embeddings.
• Evaluate unconditional vs conditional models on 4x4 up to higher resolutions, using FID and qualitative analyses including a truncation trick.

Experimental results

Research questions

• RQ1 Can meaningful, easily definable class-conditions be extracted from logos to guide generation?
• RQ2 Does conditioning enable higher-resolution logo synthesis while maintaining training stability?
• RQ3 How do different conditioning signals (object-classification vs ResNet features) affect diversity, quality, and mode coverage in logo generation?
• RQ4 What trade-offs exist between unconditional realism and conditional controllability in multi-modal logo data?

Key findings

• Unconditional StyleGAN achieves the closest match to the training distribution by FID, indicating stable, distribution-consistent output.
• Object-classification-based conditions yield diverse but less clearly separated visuals with higher FID and slower learning.
• ResNet feature-based conditions produce visually coherent and high-detail logos with strong within-class homogeneity but higher FID, indicating sharper but more divergent outputs.
• All models can train stably at a resolution four times larger than prior work, due to progressive training reducing distribution complexity at early stages.
• High-quality conditioning supports learning of more complex modes, but can introduce some nonsensical outputs and reduce global distribution fidelity.