DeSIGN: Design Inspiration from Generative Networks

Abstract: Can an algorithm create original and compelling fashion designs to serve as an inspirational assistant? To help answer this question, we design and investigate different image generation models associated with different loss functions to boost creativity in fashion generation. The dimensions of our explorations include: (i) different Generative Adversarial Networks architectures that start from noise vectors to generate fashion items, (ii) novel loss functions that encourage novelty, inspired from Sharma-Mittal divergence, a generalized mutual information measure for the widely used relative entropies such as Kullback-Leibler, and (iii) a generation process following the key elements of fashion design (disentangling shape and texture components). A key challenge of this study is the evaluation of generated designs and the retrieval of best ones, hence we put together an evaluation protocol associating automatic metrics and human experimental studies that we hope will help ease future research. We show that our proposed creativity criterion yield better overall appreciation than the one employed in Creative Adversarial Networks. In the end, about 61% of our images are thought to be created by human designers rather than by a computer while also being considered original per our human subject experiments, and our proposed loss scores the highest compared to existing losses in both novelty and likability.

• The paper introduces novel loss functions, such as Multi-class Cross Entropy and Sharma-Mittal divergence, to foster innovative fashion designs using GAN variants.
• It evaluates diverse generative models including DCGAN and StackGAN to produce high-resolution images that adhere to practical fashion constraints.
• Human assessments and adapted automatic metrics show over 60% of the designs are perceived as human-created, underscoring AI's creative potential.

Design Inspiration from Generative Networks

The paper "Design Inspiration from Generative Networks" explores the potential of artificial intelligence as an aid in the domain of fashion design. By leveraging diverse generative models, it attempts to enhance creativity in fashion generation. The study's focus encompasses both technical advancements and practical applications in AI-driven creativity.

Core Contributions

The authors propose several architectural modifications and new computational strategies to advance the generation of fashion items using GANs and their variants:

• Architectural Testing: The paper examines different generative models including DCGAN, StackGAN, and a proposed StyleGAN. These models are enhanced to generate high-resolution images that adhere to fashion design constraints, such as wearability.
• Creativity through Loss Functions: A notable contribution is the introduction of novel loss functions, specifically the Multi-class Cross Entropy (MCE) and Sharma-Mittal divergence, which encourage deviation from existing training data and foster originality in designs. These losses are compared to traditional techniques like CAN and evaluated on their ability to produce novel outputs.
• Evaluation Protocol: Given the subjective nature of creativity, the authors establish an evaluation protocol combining automatic metrics and human assessments. This multi-faceted approach helps ensure the generated designs are not only novel but also aesthetically appealing.

Methodological Insights

• Creative Adversarial Networks: Building on previous GAN frameworks, the paper extends the loss functions to promote creativity by encouraging models to produce outputs that differ from known styles and patterns.
• Disentangling Shape and Texture: A novel aspect of the research is its approach to separate shape and texture components during the generation process. This allows fashion designs to capture intricate details specific to each design element.
• Automatic Evaluation Metrics: The study introduces adapted metrics such as shape and texture scores, nearest neighbor distances, and inception-like scores, facilitating a preliminary assessment of model outputs before human evaluation.

Numerical Results and Human Evaluations

The generated fashion images were evaluated using several automatic metrics, and human subjects assessed them based on overall appreciation and originality. Notably, in certain experimental setups, over 60% of generated designs were perceived as created by human designers, indicating strong results in realism and creativity.

Implications and Future Directions

This research positions AI as a potential creative assistant in fashion design, capable of augmenting human creativity by providing inspiration and novel design alternatives. The methodologies and findings could influence broader AI applications in other creative industries by showcasing how generative models can be tailored to produce aesthetically pleasing and innovative content.

For future work, there remains potential for enhancing resolution beyond the current state and addressing remaining challenges such as the stability and reproducibility of GANs. Furthermore, long-term impacts might see these technologies integrated into design processes, offering new modes of collaboration between human designers and AI.

This paper exhibits a thoughtful exploration of creativity in AI, contributing significantly to the discourse on computational creativity and its applications in design. By articulating methodologies, results, and implications in a structured manner, it provides a promising outlook on the future of AI in creative domains.