Book Details
Steve: Framework for AI and Identity Design
Melani DeLuca
Set Margins (Eindhoven)
2024
252 pp.
ISBN: 9789083404172
Written by conversing with design studios for creating automated/augmented design software using generative AI, Steve: A Framework for AI and Identity Design, provides a rich understanding of design practice in times of rapid propagation of artificial intelligence. Steve emerges as a seminal publication for predominantly researching the question: ‘How can the most critical phases in visual identity projects be augmented through machine learning (ML)?’ Alongside adopting standard methods of design research, including design-futuring, design fiction, and collating insights from feedback sessions with designers, art directors, and clients, DeLuca’s work artistically pushes current technological limits by translating concepts of identity into collectable visual assets.
DeLuca works with visual identity understood as ‘visual elements that, together, create an atmosphere around a client, reflecting its values and views of the world and society’. These include ‘Corporate Identity,’ which stands for the set of values that an entity assumes as its own; ‘Corporate Image,’ which stands for the mental image of an entity by the public; and ‘Visual Identity,’ which stands for the representation of the corporate identity through the use of visual signs. The book’s work on design-futuring is based on Generative AI models that have increased the variety of possible prompt-based engineering-design creations and can personalise content based on a user’s preferences. The author chooses generative AI models like Stable Diffusion, MidJourney, and DALL-E in her work to create design-centric, pragmatic, custom-tailored models as per users' requirements.
De Luca's research addresses the lack of in-depth study at the intersection of graphic design theory, practice, and ML, particularly by involving practitioners in developing novel design tools. A distinctive and innovative feature of the book is its use of voice of AI in the persona of ‘Steve’. This narrative device serves a crucial purpose: to balance ‘playful writing’ with ‘rigorous academic inquiry’. In essence, it intends to equip designers with the knowledge, tools, and ethical grounding to effectively incorporate AI into their practice, ultimately advancing the methodology and practice of graphic and identity design in an increasingly AI-driven world by embodying AI and featuring testimonies from established designers.
Interviewees in the design-futuring feedback sessions context-switch between current and potential product-design strategy, including implementing AI, ML, and automated tools focusing on typography, color palette, and context. Experts providing the design technology suggest ‘instead of a full fine-tuning, it might be possible to use low-rank adaptation methods (LoRA) to optimise only a subset of model parameters’. Furthermore, DeLuca structures a set of thoughtful questions for evaluation of different aspects of practitioner feedback: ‘what is the goal of feedback?’; ‘In case of verbal/visual feedback, what kind of language / visual material is used?’; ‘How is the interaction between feedback giver and feedback receiver?’
There is research into effective prompting, vocabulary and syntax that tries to address the semantic gap to enable integration of designers’ syntax into ML datasets. Effective prompting enables the generation of artifacts that are near to the designer’s intended situational context and defines a process for inclusion of apt vocabulary and code-feature/design-parameter syntax to enable a maximized replicability of the author’s vision. The book interprets the designerly process and highlights creative moments, in particular, visual outcomes sprouting from the Definition Phase (intersection of concept construction & sketching).
By analysing a design survey (aimed at inferential, parametric, and non-parametric analysis) and corresponding ANOVA analysis on a dataset comprising designers’ syntax and verbalized design decisions, ‘the author confirms her hypothesis: graphic designers perceive designer descriptions as more useful than ML-generated ones’. The analysis is a repeated-measures ANOVA, a common inferential data analysis technique that suggests that a gradual instead of an abrupt implementation of ML would be profitable for the designers, giving way to augmentation through ML. Experts suggest ‘instead of a full fine-tuning, it might be possible to use low-rank adaptation methods (LoRA) to optimise only a subset of model parameters’.
The author ‘recommends a step-by-step implementation of ML for establishing a more tangible connection between conceptual ideas and their visual representation when creating visual identities.’ She details that systems that leverage an efficient multi-modal search combining visual and linguistic sources would be particularly advantageous for the specific requirements of design work in the Definition Phase. There is also increased focus on maintaining brand cohesion over time using a robust technical implementation via PyTorch for apt model building, training, and testing. The designers’ approach—conceptualization, building, and testing—enables others to be informed about performance and make necessary adjustments.
Datasets contain rich attribute labels for content, emotions, and artistic media. Platforms such as It’s Nice That, Behance, Pinterest, and The Typo/Graphic Poster archive are researched for image typologies, perception of compositions and typographical variations, type sizes, file compression, and colour values. DeLuca recommends integrating more graphic and typographic materials into ML datasets and subsequently fine-tuning to incorporate increased semantic cognition of design, ‘enabling a proficient creative process to curate small datasets and tag visual assets using customized semantics and perspectives with interactive ML training’.
Coining it ‘Steve,’ ‘the framework for augmented visual identity design process,’ shows similarity with multimodal contrastive learning, called CLIP (Contrastive Language Image Pre-training), consisting of visual and textual modalities – given an image, the model can predict the most relevant text description for that image and analyse its content in relation to various artists, media and styles. The architecture of Steve is inherently shaped by technical choices, e.g., ‘a contrastive loss function for learning cross-modal embeddings, an adaptive learning rate such as Adam or AdamW to improve the training rates and integration of GPT agent clusters’. With the objective of profoundly shaping public perception and societal narratives, the conceptual framework is designed to integrate the designer's perspective into the development and use of new digital tools. A broader implication is that Steve is not just a guide for how to use AI, but a call for responsible and culturally aware AI applications.
In conclusion, the framework avails a process of empowered designing for creators. Steve: A Framework for AI and Identity Design is more than just a manual; it is a foundational text that initiates a critical dialogue and provides a replicable conceptual structure for designers and researchers. It is a book that explores how AI and ML can be ethically and effectively integrated into graphic design, particularly in the realm of identity design. It is an indispensable resource for responsibly navigating the evolving intersection of AI and creativity, shaping the future of identity design.
Maithili Mishra is a designer and mentor from Clemson University. A budding design-engineering researcher, she specializes in the research areas of “Human-Centered Computing (HCC),” specifically “virtual humans (VH)” and is working on “KIMONO (keeping it monochromatic).” She uses Generative AI applications, and is interested in the pragmatic development, critique, evaluation, and characterization of UX and visualization systems (topology, form, and color). She enjoys embodying a connoisseur’s outlook that brings up nuanced details of articulate representativeness. She can be found on Medium and X on handles https://medium.com/@maithilim47 and https://x.com/methyl_orange47