Integrating AIGC with design: dependence, application, and evolution - a systematic literature review
将 AIGC 与设计整合：依赖性、应用和演变 - 一项系统文献综述

3.1.1. Search and identification
3.1.1. 搜索和识别

The study used the Scopus database as the primary source for the literature search, which is the largest peer-reviewed abstract and citation database covering the indexed content of numerous journals and conferences from literature databases such as Science Direct, the Sage, ACM Digital Library, IEEE Xplore Digital Library, Taylor & Francis Online, and others. The literature review will incorporate peer-reviewed journals and high-quality conference articles, considering that AIGC is at the forefront of content in the AI research area and that many high-quality studies are published at the conference.
该研究使用 Scopus 数据库作为文献检索的主要来源，Scopus 是最大的同行评议摘要和引文数据库，涵盖了来自诸如 Science Direct、Sage、ACM 数字图书馆、IEEE Xplore 数字图书馆、Taylor & Francis Online 等文献数据库的众多期刊和会议的索引内容。文献综述将包括同行评议的期刊和高质量的会议文章，考虑到 AIGC 在人工智能研究领域处于领先地位，并且许多高质量的研究都是在会议上发表的。

This study aims to capture research themes and paradigms from the literature on the application of AIGC in design. Consequently, term combinations related to both domains (AIGC and design) were employed to identify peer-reviewed articles. The literature retrieval process was conducted using two dimensions: thematic term searches and technical multidimensional term searches. In terms of thematic term searches, given that the investigation of this study revolves around the application of AIGC in the design domain, AIGC was employed as the primary term. This encompassed related terms such as ‘Artificial Intelligence Generated Content,’ ‘AIGC,’ ‘Generative Artificial Intelligence,’ ‘Generative AI,’ and ‘GenAI.’ In terms of technical multidimensional term searches, considering that the proprietary terms for AIGC gradually gained popularity after 2022, while generative technologies have been developed and widely utilised since 2014, the technical dimensions of generative artificial intelligence were also included in the search query. These encompassed terms such as ‘Generative Adversarial Network,’ ‘GAN,’ ‘Variational Auto-Encoder,’ ‘VAE,’ ‘Diffusion Model,’ ‘Neural Network,’ ‘Natural Language Processing,’ ‘NLP,’ ‘Deep Learning,’ and ‘DL.’ By using ‘Generat*,’ the search content was restricted specifically to the generative domain. More literature in Architecture, Chemistry, and Pharmaceuticals were found in the search results and excluded using ‘Not’; the detailed search formula is shown in Table 
本研究旨在从应用 AIGC 于设计领域的文献中捕捉研究主题和范式。因此，采用与两个领域（AIGC 和设计）相关的术语组合来识别同行评议的文章。文献检索过程采用了两个维度：主题术语搜索和技术多维术语搜索。在主题术语搜索方面，考虑到本研究的调查围绕 AIGC 在设计领域的应用展开，AIGC 被用作主要术语。这包括相关术语，如“人工智能生成内容”，“AIGC”，“生成人工智能”，“生成 AI”和“GenAI”。在技术多维术语搜索方面，考虑到 AIGC 的专有术语在 2022 年后逐渐流行起来，而生成技术自 2014 年以来已经得到发展和广泛应用，生成人工智能的技术维度也被包括在搜索查询中。 这些术语包括“生成对抗网络”，“GAN”，“变分自动编码器”，“VAE”，“扩散模型”，“神经网络”，“自然语言处理”，“NLP”，“深度学习”和“DL”。通过使用“Generat*”，搜索内容被限制在生成领域。在搜索结果中发现了更多的建筑学、化学和制药学文献，并使用“不”进行排除；详细的搜索公式如表所示1. The last search was conducted in December 2023, and the initial search yielded a total of 528 entries.
最后一次搜索是在 2023 年 12 月进行的，最初的搜索结果总共有 528 个条目。

3.1.2. Screening, eligibility, and inclusion
3.1.2. 筛选、资格和纳入

To ensure the rigour and credibility of the included literature, the study established stringent inclusion and exclusion criteria, subjected to three independent reviews at different levels by two expert assessors. The assessors demonstrated excellent consistency during the review process (Cohen’s Kappa > 0.8), resolving divergent cases through discussions to achieve consensus. The detailed review steps are as follows:
为确保所包含文献的严谨性和可信度，研究建立了严格的纳入和排除标准，由两名专家评估者在不同层次进行三次独立审查。评估者在审查过程中表现出优秀的一致性（Cohen's Kappa > 0.8），通过讨论解决分歧案例以达成共识。具体审查步骤如下：

First, 528 documents obtained from the search were initially reviewed, and duplicates with the same article title, content, and authors, non-peer-reviewed documents, non-English documents, and documents for which the full text was not available were excluded, totalling 19 documents.
首先，对从搜索中获得的 528 份文件进行了初步审阅，排除了与相同文章标题、内容和作者重复的文件、非同行评议文件、非英文文件以及全文不可用的文件，共计 19 份文件。

Further, a literature content review was conducted based on inclusion and exclusion criteria, and only research literature at the intersection of the two fields of AIGC and design was included. The specific criteria are outlined in Table 
此外，根据纳入和排除标准进行了文献内容审查，仅包括在 AIGC 和设计两个领域交叉点上的研究文献。具体标准详见表中。2. The second review was conducted by reading the titles and abstracts, and 318 non-compliant documents were excluded. The third review was conducted by traversing the complete text, and 83 documents that did not meet the criteria were excluded. Finally, a total of 90 qualified documents were included in the corpus, of which 48 were from journals, and 42 from conferences, and the details of the literature are shown in
第二次审查是通过阅读标题和摘要进行的，排除了 318 份不符合要求的文件。第三次审查是通过遍历完整文本进行的，排除了 83 份不符合标准的文件。最终，共有 90 份合格文件被纳入语料库，其中 48 份来自期刊，42 份来自会议，文献的详细信息如下。Tables A1 表 A1 and  和A2 in the Appendix.
在附录中。

5.3.1. Creative divergence
5.3.1. 创意分歧

In the early stages of design conceptualisation, designers must integrate their internal knowledge and a broad range of external expertise to seek solutions to problems. However, due to individual cognitive biases, there is often a tendency for design fixation, indicating that designers are inclined to adhere to pre-existing ideas or concepts, ultimately restricting the design outcome. AIGC has been effective in bringing inspiration to designers by developing Creative Support Tools (CSTs) due to its extensive database base and diversity of generated content. It primarily involves three aspects (Table 
在设计概念化的早期阶段，设计师必须整合他们的内部知识和广泛的外部专业知识，以寻求解决问题的方案。然而，由于个体认知偏见的存在，往往存在设计固化的倾向，表明设计师倾向于坚持现有的想法或概念，最终限制了设计结果。AIGC 通过开发创意支持工具（CSTs）为设计师带来灵感，因为它具有庞大的数据库和多样化的生成内容。它主要涉及三个方面（表格）。6): providing visual and semantic stimuli through analogical design, extracting design insights by mining user needs, and exploring interactive design space.
通过模拟设计提供视觉和语义刺激，通过挖掘用户需求提取设计洞察，探索交互设计空间。

Providing visual and semantic stimuli through analogical design. Design-by-Analogy (DbA) is an ideation strategy that seeks inspiration from a source domain and generates design concepts in a target domain. The human mind is associative, and providing relevant analogical inspiration can activate the designer's long-term memory and facilitate design conceptualisation. AIGC assists designers in creative divergence through stimulus retrieval and stimulus generation.
通过类比设计提供视觉和语义刺激。类比设计（DbA）是一种构思策略，它从源领域寻找灵感，并在目标领域生成设计概念。人类的思维是联想的，提供相关的类比灵感可以激活设计师的长期记忆，并促进设计概念的形成。AIGC 通过刺激检索和刺激生成帮助设计师进行创意分歧。

• Stimulus retrieval. By recognising and clustering the content in the corresponding text or image databases, the distance between the design goal and the database content is measured in multiple dimensions, which helps designers quickly retrieve and locate similar concepts, thus facilitating inspirational inspiration. Semantic stimulus retrieval mainly relies on natural language processing techniques (NLP) to understand the contextual information of the corpus through modelling algorithms such as word2vec, Long Short-Term Memory Network (LSTM), and Transformer, which represent the textual information in a vector space (Li et al. Citation2024). Database sources include design logs, patent repositories, product documentation, etc. e.g. Liu (Citation2020) used patent databases as a source of analogical knowledge, constructed content relevance measurement dimensions based on structure mapping theory (SMT), and realised the search for design analogies to provide designers with relevant knowledge and information insights. Visual stimulus retrieval relies on deep learning to recognise and cluster visual similarity and conceptual similarity of images. Karimi et al. (Citation2019) is based on Quickdraw's library of hand-drawn sketches and utilises Convolutional Neural Networks (CNNs) to learn potential shape features of multiple sketch categories and perform clustering. As designers sketch, the system can present patterns that are visually or conceptually relevant to the user's input sketch to stimulate design creativity.
  刺激检索。通过识别和聚类相应文本或图像数据库中的内容，测量设计目标与数据库内容之间在多个维度上的距离，帮助设计师快速检索和定位类似概念，从而促进灵感的产生。语义刺激检索主要依赖自然语言处理技术（NLP）来通过建模算法如 word2vec、长短期记忆网络（LSTM）和 Transformer 理解语料库的上下文信息，这些算法将文本信息表示在一个向量空间中（Li 等，2024）。数据库来源包括设计日志、专利库、产品文档等。例如，Liu（2020）利用专利数据库作为类比知识的来源，基于结构映射理论（SMT）构建内容相关性测量维度，并实现了对设计类比的搜索，为设计师提供相关知识和信息见解。视觉刺激检索依赖深度学习来识别和聚类图像的视觉相似性和概念相似性。Karimi 等。 2019 年基于 Quickdraw 的手绘草图库，利用卷积神经网络（CNNs）学习多种草图类别的潜在形状特征并进行聚类。设计师在草绘时，系统可以呈现与用户输入草图在视觉或概念上相关的图案，以激发设计创意。

• Stimulate generation. Focuses primarily on the visual domain and relies on image style migration techniques. By capturing conceptual elements from different domains and reorganising them, AIGC can integrate imagery elements with original designs to generate unique and novel content that stimulates designers’ combinatorial creativity (Wang et al. Citation2021). This type of style and feature migration is more often used in clothing and textile pattern design. e.g. Padiyath and Magerko (Citation2021) proposed a ‘bad’ Generative Adversarial Network (GAN) that integrates surreal images and clothing patterns to generate unique and bizarre clothing images to inspire designers. Yan (Citation2023) constructs a generative model for ‘inspiration’ transfer, which realises the transfer of material properties from the source fashion item to the target design.
  刺激生成。主要关注视觉领域，并依赖于图像风格迁移技术。通过捕捉不同领域的概念元素并重新组织它们，AIGC 可以将图像元素与原始设计整合在一起，生成刺激设计师组合创造力的独特和新颖内容（Wang 等，2021）。这种风格和特征迁移更常用于服装和纺织图案设计。例如，Padiyath 和 Magerko（2021）提出了一个“坏”的生成对抗网络（GAN），将超现实图像和服装图案整合在一起，生成独特和奇异的服装图像，以激发设计师的灵感。Yan（2023）构建了一个“灵感”转移的生成模型，实现了将源时尚物品的材料属性转移到目标设计中。

Extracting design insights by mining user needs. At the early stage of design, it is essential to consider users’ needs and preferences, which helps to clarify the design positioning and strategy. Traditional user research methods rely on questionnaires and interviews, which are time-consuming and cumbersome, with a low data reuse rate. AIGC relies on powerful data processing capabilities to provide new methods for user needs mining.
通过挖掘用户需求来提取设计洞见。在设计的早期阶段，考虑用户的需求和偏好至关重要，这有助于澄清设计定位和策略。传统的用户研究方法依赖于问卷调查和访谈，这些方法耗时且繁琐，数据重复利用率低。人工智能图形计算依赖于强大的数据处理能力，为用户需求挖掘提供了新方法。

• Mining requirements from text. Users’ evaluation and feedback on products contain personal needs and preferences, and the user evaluation system of online shopping platforms provides an excellent channel for collecting user feedback. By identifying users’ opinion words and sentiment words in online reviews through natural language processing (NLP) and constructing a demand model of product attributes and consumer emotions based on sentiment analysis (SA), it can achieve the importance weighting metric of product attributes and the inference of users’ sentiments, and provide optimisation strategies for product improvement (Saidani, Kim, and Yannou Citation2021).
  从文本中挖掘需求。用户对产品的评价和反馈包含个人需求和偏好，而在线购物平台的用户评价系统为收集用户反馈提供了一个极好的渠道。通过通过自然语言处理（NLP）在在线评论中识别用户的意见词和情感词，并基于情感分析（SA）构建产品属性和消费者情绪的需求模型，可以实现产品属性的重要性加权度量和用户情感的推断，并为产品改进提供优化策略（Saidani，Kim 和 Yannou 2021）。

• Mapping of emotional preferences. Kansei Engineering, as a design method to quantitatively analyze the emotion of products, is combined with AIGC to quickly generate design concepts that satisfy users’ aesthetic and emotional needs. Firstly, a rich dataset is constructed which includes a large number of images and multi-dimensional emotion labelling. Train deep learning models to realise the recognition and classification of image emotions, and establish the mapping relationship between design attributes, styles and user preferences. Further, construct a generative model to merge and reconstruct features such as colours and attributes that match specific emotions with the target product by means of neural style migration to achieve the generation of design concepts that match user preferences (Quan, Li, and Hu Citation2018; Li et al. Citation2021).
  情感偏好的映射。将感性工程作为一种设计方法，与 AIGC 相结合，快速生成满足用户审美和情感需求的设计概念。首先，构建一个包括大量图像和多维情感标记的丰富数据集。训练深度学习模型实现图像情感的识别和分类，并建立设计属性、风格和用户偏好之间的映射关系。进一步，通过神经风格迁移的方式构建生成模型，合并和重构与目标产品匹配特定情感的特征，如颜色和属性，以实现生成符合用户偏好的设计概念（Quan, Li, and Hu 2018; Li 等，2021）。

• Incorporating physiological signals. Physiological signals are associated and mapped to user preference requirements. The researchers propose a neurocognitive-inspired machine learning and design generation approach. Human cognitive factors are measured by EEG to intervene in the image generation process, thus realising the design of image generation in accordance with user preferences (Wang et al. Citation2020).
  将生理信号纳入考虑。生理信号与用户偏好需求相关联并映射。研究人员提出了一种神经认知启发的机器学习和设计生成方法。通过脑电图测量人类认知因素，干预图像生成过程，从而实现根据用户偏好生成图像的设计（Wang 等，2020 年）。

Exploring interactive design space. Creative ideation is an evolutionary and iterative process. Basic generative tools tend to provide only one-time outputs that lack controllability and are not iterative. AIGC will become a more effective creativity support tool based on design thinking to enable continuous design space exploration and improved interactive communication performance with designers.
探索互动设计空间。创意构思是一个进化和迭代的过程。基本生成工具往往只提供一次性输出，缺乏可控性且不具迭代性。基于设计思维的 AIGC 将成为一种更有效的创意支持工具，以实现持续的设计空间探索，并提高与设计师的互动沟通表现。

• Spatial exploration based on design thinking. Creativity involves divergent and convergent thinking, encompassing the generation of a multitude of creative ideas and the refinement of valuable concepts. The continuous iteration of creativity within the realms of divergence and convergence can be perceived as an ongoing design space exploration. Davis et al. (Citation2023) constructed a potential design space exploration panel based on generative models by analyzing design divergent and convergent thinking, allowing users to move the image via the X and Y axes to generate continuous and controllable difference images to assist designers in creative conceptualisation.
  基于设计思维的空间探索。创造力涉及发散和收敛思维，包括产生大量创意想法和精炼有价值概念。在发散和收敛的领域内持续迭代创造力可以被视为持续的设计空间探索。Davis 等人（2023 年）通过分析设计的发散和收敛思维，构建了一个潜在的基于生成模型的设计空间探索面板，允许用户通过 X 和 Y 轴移动图像，生成连续可控的差异图像，以帮助设计师进行创意概念化。

• Ideation collaboration based on interactive communication. To break through the limitation of the unidirectional output of traditional generative models, a system of bidirectional communication and multistep collaboration between AIGC and human designers needs to be constructed urgently. Ibarrola, Lawton, and Grace (Citation2023) develops a co-creation drawing system that senses the semantics of the designer's sketches. AIGC and the designer interact collaboratively, negotiating and iterating on design ideas.
  基于互动沟通的构思协作。为了突破传统生成模型单向输出的局限，迫切需要建立 AIGC 与人类设计师之间的双向沟通和多步协作系统。Ibarrola、Lawton 和 Grace（2023）开发了一种共创绘图系统，能感知设计师草图的语义。AIGC 和设计师之间进行互动协作，就设计理念进行协商和迭代。

Creative divergence is a complex process to solve undefined or unknown problems, it is necessary to pay attention to user demand preferences as well as consider the novelty and creativity of ideas. However, the use of a single text-to-image tool in the ideation process leads instead to more severe initial fixation. The time and effort expended on the prompt project led to sunk costs, and overly detailed generation outcomes result in conscious fixation by designers (Wadinambiarachchi et al. Citation2024). Therefore, it is crucial to develop a reasonable AIGC creativity support system based on the characteristics of design conception: optimising the path of data mining and user demand analysis to enhance the breadth and efficiency of insights; relying on analogical design and design thinking theory to propose strategies for AIGC-assisted conception. This type of application is moving towards more diversified and controllable.
创意分歧是解决未定义或未知问题的复杂过程，需要关注用户需求偏好，并考虑想法的新颖性和创造性。然而，在构思过程中仅使用单一的文本到图像工具反而导致更严重的初始固定化。在快速项目上花费的时间和精力导致了沉没成本，过于详细的生成结果导致设计师的意识固定化（Wadinambiarachchi 等，2024 年）。因此，基于设计构思特点，开发一个合理的 AIGC 创意支持系统至关重要：优化数据挖掘和用户需求分析路径，以增强洞察力的广度和效率；依靠类比设计和设计思维理论，提出 AIGC 辅助构思的策略。这种应用正在朝着更多样化和可控性的方向发展。

5.3.2. Design generation 5.3.2. 设计生成

AIGC is widely used in the design generation phase. The design domain has been following specific steps to create new products, but in the face of the vast consumer market and constantly updated user needs, automating the design work using computer technology is beneficial to improve design efficiency, reduce cost, and increase efficiency. AIGC can be classified into a generation tool for labour-intensive tasks and a design assistant for human–computer collaboration according to its job definition in the design generation task (
AIGC 广泛应用于设计生成阶段。设计领域一直遵循特定步骤来创建新产品，但面对庞大的消费市场和不断更新的用户需求，利用计算机技术自动化设计工作有助于提高设计效率，降低成本，提高效率。根据其在设计生成任务中的工作定义，AIGC 可以被分类为劳动密集型任务的生成工具和人机协作的设计助手。Table 7 表 7).

The generation tool for labour-intensive tasks. Automated generation tools have the capability to produce a large number of design images with consistent forms and compliance with specified requirements in an unsupervised environment, enabling batch mass production. Therefore, it is valuable in the fast production and fast sales industry where there is a huge demand for design, such as textile pattern generation and graphic logo drawing (Yang et al. Citation2021; Mao, Wang, and Jiang Citation2020). It is also effective in repetitive design work, such as in the field of typeface design, where a small number of source glyphs are used as input to complement other fonts with consistent morphological characteristics (Yan et al. Citation2020), and in assisted sketching for colouring, where it generates uniform and coherent sketches from multiple perspectives (Jiang, Li, and Wang Citation2021). This kind of AIGC tool has a unified construction logic: first build a dataset oriented to a specific design task, and further train and generate models based on deep learning algorithms such as GAN and VAE. Different design tasks require the model to perceive and learn different features, so the algorithms need to be optimised and fine-tuned on the established framework (Figure 
劳动密集型任务的生成工具。自动生成工具具有在无监督环境中生成大量设计图像的能力，形式一致且符合指定要求，从而实现批量大规模生产。因此，在快速生产和快速销售行业中具有价值，这些行业对设计有巨大需求，如纺织图案生成和图形标志绘制（杨等，2021 年；毛，王和江，2020 年）。它还在重复设计工作中发挥作用，比如在字体设计领域，使用少量源字形作为输入，以补充其他具有一致形态特征的字体（严等，2020 年），以及在辅助素描上色中，从多个角度生成统一和连贯的素描（江，李和王，2021 年）。这种 AIGC 工具具有统一的构建逻辑：首先构建面向特定设计任务的数据集，然后基于深度学习算法（如 GAN 和 VAE）进行进一步训练和生成模型。不同的设计任务需要模型感知和学习不同的特征，因此算法需要在已建立的框架上进行优化和微调（见图）6). For example, in the design generation task of apparel patterns, local texture details are handled by incorporating a multi-scale discriminator in the model and a self-attention mechanism is introduced to improve global art perception (Yu and Luo Citation2021). In product generation, a Structured Deformable Meshes Net (SDM-NET) is used to construct a perception module for complex geometric shapes (Li, Xie, and Sha Citation2023).
例如，在服装图案设计生成任务中，通过在模型中加入多尺度鉴别器来处理局部纹理细节，并引入自注意机制来提高全局艺术感知（Yu 和 Luo 2021）。在产品生成中，使用结构可变网格网络（SDM-NET）构建感知模块，用于复杂几何形状（Li，Xie 和 Sha 2023）。

Figure 6. AIGC construction differences for different design objects.
图 6. 不同设计对象的 AIGC 构建差异。

Display full size

The design assistant for human–computer collaboration. Creative design tasks involve multiple steps and processes. It is important to build a human-machine collaboration system based on design cognition. On the one hand, dismantling the design process allows for collaboration on tasks in which the machine and the designer have their specialties. On the other hand, integrating computational thinking and design thinking enables intelligent generation that depends on design cognition.
人机协作的设计助手。创意设计任务涉及多个步骤和过程。基于设计认知构建人机协作系统至关重要。一方面，拆解设计过程允许机器和设计师在各自擅长的任务上进行协作。另一方面，整合计算思维和设计思维实现依赖设计认知的智能生成。

• Deconstruction of the design process. AIGC and designers each have specialised capabilities. By disassembling the design process, a collaborative system can be constructed that is generated by AIGC and decided by designers. For example, Wu (Citation2023) disassembles the apparel design process and builds a human–computer collaboration system: StyleMe. In the sketch generation stage, the system assists in generating rough line drawings based on the styles selected by the designer, and the designer makes fine-tuning adjustments. Furthermore, the system performs style transfer based on the designer's selected thematic reference images to generate realistic apparel images.
  设计过程的解构。AIGC 和设计师各自拥有专业能力。通过解构设计过程，可以建立一个由 AIGC 生成并由设计师决定的协作系统。例如，吴（2023）解构服装设计过程并构建了一个人机协作系统：StyleMe。在草图生成阶段，该系统协助根据设计师选择的风格生成粗略线描，并由设计师进行微调。此外，该系统根据设计师选择的主题参考图像执行风格转移，生成逼真的服装图像。

• Incorporation of design cognition. Relying on the guidance of design theory, incorporating design cognition into the construction of generative models is an essential trend in integrating AIGC and design. Wu (Citation2023) established a cognitive mental model of automobile frontal styling from the three dimensions of design intent, drawing behaviour, and functional structure, constructed a human-machine shared knowledge base through image acquisition, hand-drawn sketches, and morpho-semantic annotation, and constructed a conditional cross-domain generative adversarial network information model by establishing a morpho-semantic encoding network, a sketching feature extraction network, and an image-perception clustering network, which realised the mapping from human cognitive space to computational space to realise the creative generation of automobile frontal styling based on sketch-semantic mapping. This human–computer collaborative generative system gives the user a high degree of control over the design process. It accurately guides the AI to create the design according to the target direction.
  设计认知的整合。依靠设计理论的指导，将设计认知纳入生成模型的构建是整合 AIGC 和设计的一个重要趋势。吴（2023）从设计意图、绘图行为和功能结构三个维度建立了汽车前部造型的认知心理模型，通过图像获取、手绘草图和形态语义标注构建了人机共享知识库，并通过建立形态语义编码网络、草图特征提取网络和图像感知聚类网络构建了条件跨领域生成对抗网络信息模型，实现了从人类认知空间到计算空间的映射，实现了基于草图语义映射的汽车前部造型的创造性生成。这种人机协作生成系统赋予用户对设计过程的高度控制。它准确引导 AI 根据目标方向创建设计。

Design generation is the most direct and widespread application of AIGC as a content output tool. AIGC builds applications based on the characteristics and output requirements of different design tasks, effectively speeding up the design process and increasing the output, easing the pressure of labour-intensive work in design. In addition, human-machine collaborative AIGC design systems are rapidly emerging, generating models to produce according to instructions, with designers responsible for decision-making and deployment. This type of application has gradually demonstrated the ability to rival or even surpass traditional professional designers. This is further evidence that the basic generative capabilities of AIGC can be optimised and constructed to handle the complex activity of design.
设计生成是 AIGC 作为内容输出工具的最直接和普遍的应用。AIGC 根据不同设计任务的特点和输出要求构建应用程序，有效地加快设计过程并增加输出，减轻设计中劳动密集型工作的压力。此外，人机协作的 AIGC 设计系统正在迅速出现，生成模型以根据指令进行生产，设计师负责决策和部署。这种应用逐渐展示出与传统专业设计师相媲美甚至超越的能力。这进一步证明了 AIGC 的基本生成能力可以被优化和构建以处理设计的复杂活动。

5.3.3. Assistance and advice
5.3.3. 协助和建议

AIGC can also act as a ‘consultant’ in the design process, providing designers with design assistance advice. Instead of generating target designs directly, these AIGC applications analyze design data and trends, focusing on the designer's preferences and project requirements. Further, provide designers with stylistic advice on layout, colour, materials, etc., and assist guidance on the design process (
AIGC 还可以在设计过程中充当“顾问”，为设计师提供设计辅助建议。这些 AIGC 应用程序不直接生成目标设计，而是分析设计数据和趋势，关注设计师的偏好和项目要求。此外，为设计师提供关于布局、颜色、材料等方面的风格建议，并在设计过程中提供指导。Table 8 表 8).

Stylised recommendations. The realisation of design concepts follows specific aesthetic logic and rules, and AIGC provides standardised style suggestions to assist designers in reaching their design goals by learning from existing design cases.
风格化建议。设计概念的实现遵循特定的美学逻辑和规则，AIGC 提供标准化的风格建议，以帮助设计师通过借鉴现有的设计案例来实现他们的设计目标。

• Layout generation and optimization. Assistive suggestions for design layout are mainly focused on areas such as graphic design and UI design that involve typographic aesthetics. For example, in graphic typographic layout, saliency prediction models are trained to perceive blank areas of an image, find appropriate text placement, and evaluate the aesthetic quality of candidate results for output (Zhang, Li, and Wang Citation2020). A key goal of graphical user interface (GUI) design is to ensure that the same product remains consistent across screens. Brückner, Leiva, and Oulasvirta (Citation2022) constructed a design-assist plug-in built into Sketch based on the visual properties of interfaces, spatial patterns, and repetitive patterns of group elements to provide layout consistency recommendations when designers add elements.
  布局生成与优化。设计布局的辅助建议主要集中在涉及排版美学的领域，如平面设计和 UI 设计。例如，在平面排版设计中，显著性预测模型被训练用于感知图像的空白区域，找到适当的文本放置位置，并评估候选结果的美学质量以输出（张、李和王，2020）。图形用户界面（GUI）设计的一个关键目标是确保同一产品在不同屏幕上保持一致。Brückner、Leiva 和 Oulasvirta（2022）基于界面的视觉属性、空间模式和组元素的重复模式构建了一个设计辅助插件，用于在设计师添加元素时提供布局一致性建议。

• Color and material suggestions. Colour coordination is a crucial aspect of design, but achieving optimal colour schemes demands extensive design experience and a wealth of colour-related knowledge. AIGC relies on a wide range of colour-matching datasets to train deep learning models, enabling the generation of appropriate colour-matching schemes based on different input themes (Feng et al. Citation2021). Under the guidance of Kansei Engineering, Ding and Dong (Citation2019) proposed a product colour emotional design method, which constructs a mapping model between user colour perception and product design elements through neural networks, and recommends colour layouts for designers to match the user's emotional preferences. AIGC-assisted colour matching generation has been widely used in product design ranks, such as automated stylised rendering of automobile side profiles (Ji and Chen Citation2022), etc.
  颜色和材料建议。颜色协调是设计的关键方面，但要实现最佳的配色方案需要丰富的设计经验和丰富的与颜色相关的知识。AIGC 依赖于各种颜色匹配数据集来训练深度学习模型，从而能够根据不同的输入主题生成适当的配色方案（Feng 等，2021 年）。在感性工程的指导下，丁和董（2019 年）提出了一种产品颜色情感设计方法，通过神经网络构建用户颜色感知与产品设计元素之间的映射模型，并推荐设计师匹配用户情感偏好的颜色布局。AIGC 辅助的颜色匹配生成已广泛应用于产品设计领域，例如汽车侧面轮廓的自动化风格化渲染（季和陈，2022 年）等。

Steps and process guidance. Large Language Models (LLMs) are constructed based on billions of data and belong to the generalised domain, e.g. GPT-3, ChatGPT. Prefabricated design steps in LLM to build custom small models which make it steer towards design tasks. The designer interacts and communicates with AIGC with specific prompts to get design guidance. For example, Torii, Murakami, and Ochiai (Citation2023) investigates an approach to board game creation called ‘Lottery and Sprint’ that combines human design intuition with a structured design sprint framework executed by the AutoGPT system, aiming to facilitate a collaborative game design experience by aligning an AI-driven process with human creativity. This represents an indirect application approach. Consequently, the construction process entails inherent challenges and uncertainties. In the future, it is imperative to develop controlled, process-assistive applications tailored explicitly for professional design domains.
步骤和流程指导。大型语言模型（LLMs）是基于数十亿数据构建的，属于泛化领域，例如 GPT-3，ChatGPT。在LLM中预制设计步骤用于构建定制的小型模型，使其朝向设计任务。设计师通过特定提示与 AIGC 互动和沟通，以获得设计指导。例如，Torii，Murakami 和 Ochiai（2023）研究了一种名为“抽签和冲刺”的棋盘游戏创作方法，该方法将人类设计直觉与 AutoGPT 系统执行的结构化设计冲刺框架相结合，旨在通过将 AI 驱动的流程与人类创造力对齐，促进协作游戏设计体验。这代表了一种间接的应用方法。因此，构建过程涉及固有的挑战和不确定性。未来，开发专门针对专业设计领域的受控、过程辅助应用是至关重要的。

5.3.4. Evaluation and feedback
5.3.4. 评估与反馈

In the traditional design process, products must be manufactured and used by consumers before evaluations and feedback can be gathered, which introduces a delay. AIGC, relying on deep learning and neural networks, is capable of accurately uncovering patterns within large datasets and making precise predictions. It enables the acquisition of prior evaluations and feedback, and can even perform automatic optimizations. This significantly enhances the efficiency of design iterations and reduces trial-and-error costs. Existing concerns about the application of AIGC to design evaluation and feedback processes cover various aspects, such as similarity detection for anti-plagiarism, advanced verification of design novelty, and feasibility feedback for cost and quality (
在传统的设计过程中，产品必须在制造和消费者使用之后才能收集评估和反馈，这会引入延迟。依靠深度学习和神经网络，AIGC 能够准确地揭示大型数据集中的模式并进行精确预测。它使得先前的评估和反馈的获取成为可能，甚至可以执行自动优化。这显著提高了设计迭代的效率并降低了试错成本。关于将 AIGC 应用于设计评估和反馈过程的现有关注点涵盖了各个方面，如反抄袭的相似性检测、设计新颖性的高级验证以及成本和质量的可行性反馈。Table 9 表 9).

• Similarity detection for anti-plagiarism. Copyright awareness of design is very important in the feasibility evaluation of design, especially in the fields of graphic design and graphic design. The plagiarism features are extracted and trained in a deep learning model to achieve the recognition of similar content such as pattern and layout. Liu (2023) proposed a visual saliency feature extraction network based on the GAN model, which captures global high-level semantic information of graphic design, refines elemental relational structure via image segmentation and performs multi-weight similarity measurement. Yang (2021) proposes a method for measuring the perceived similarity of graphics that is capable of detecting different types of logo plagiarism including duplicates, gradients, superimpositions, perspectives, and three-dimensional deformations.
  反抄袭的相似性检测。版权意识在设计的可行性评估中非常重要，特别是在平面设计和图形设计领域。通过提取和训练深度学习模型中的抄袭特征，实现对图案和布局等相似内容的识别。刘（2023）提出了基于 GAN 模型的视觉显著性特征提取网络，捕捉图形设计的全局高级语义信息，通过图像分割细化元素关系结构，并进行多权重相似度测量。杨（2021）提出了一种测量图形感知相似性的方法，能够检测不同类型的标志抄袭，包括重复、渐变、叠加、透视和三维变形。

• Advance verification of design novelty. Novelty evaluation and aesthetic assessment are essential aspects of generative models for design evaluation which rely on neural networks to construct a mapping model of design features and aesthetic evaluation. Heyrani Nobari (2021) proposed an automatic method for generating novel designs, CreativeGAN, which can measure the overall product novelty by extracting features and using the K-nearest neighbour algorithm, locate critical components and their attributes affecting the novelty, and further automatically modify the GAN to achieve directed generation of novel components. In interface design, Schoop et al. (Citation2022) uses a visual modelling and saliency analysis approach to predict and interpret the clickability of mobile UI through deep learning, providing designers with insights to evaluate and improve their designs quickly.
  设计新颖性的提前验证。新颖性评估和审美评估是设计评估生成模型的基本方面，依赖于神经网络构建设计特征和审美评估的映射模型。Heyrani Nobari（2021）提出了一种用于生成新颖设计的自动方法 CreativeGAN，可以通过提取特征并使用 K 最近邻算法来衡量整体产品的新颖性，定位影响新颖性的关键组件及其属性，并进一步自动修改 GAN 以实现新颖组件的有向生成。在界面设计方面，Schoop 等人（2022）采用视觉建模和显著性分析方法通过深度学习来预测和解释移动 UI 的可点击性，为设计师提供洞察力，以快速评估和改进他们的设计。

• Feasibility feedback on cost and quality. The insight and feedback of information are mainly achieved through LLM, which relies on an extensive database. Gallega and Sumi (Citation2023) proposed a co-creation system that utilises AIGC to assist designers in changing and rendering 3D materials and accessing the feasibility and cost analysis of the materials plugging into GPT-3.
  成本和质量的可行性反馈。信息的洞察和反馈主要通过LLM实现，该系统依赖于庞大的数据库。加列加和苏米（2023）提出了一个共创系统，利用 AIGC 来帮助设计师改变和渲染 3D 材料，并访问材料插入 GPT-3 的可行性和成本分析。

Figure 7. AIGC-aided design workflow.
图 7. AIGC 辅助设计工作流程。

Display full size

6.1.1. From point-to-point goal generation to controlled design space exploration
6.1.1. 从点对点目标生成到受控设计空间探索

Given the inherent uncertainty of design problems and the non-uniqueness of solutions, Simon (Citation1973) characterised design as ill-structured problems. Designers are required to identify and define problems through insights and utilise creativity to develop diverse solutions. This process is inherently exploratory, involving continuous experimentation and iterative refinement. Deep generative models map various raw data features onto latent spaces, making these continuous data features are especially suitable for supporting exploration in the design space (Burnap et al. Citation2016).
鉴于设计问题的固有不确定性和解决方案的非唯一性，西蒙（1973）将设计描述为结构不清晰的问题。设计师需要通过洞察力识别和定义问题，并利用创造力开发多样化的解决方案。这个过程本质上是探索性的，涉及持续的实验和迭代的完善。深度生成模型将各种原始数据特征映射到潜在空间，使得这些连续数据特征特别适合支持设计空间中的探索（Burnap 等人，2016）。

Since the inception of generating model construction, there has been continuous optimisation focused on enhancing the quality and efficiency of generation within computational intelligence. In its initial applications within the design domain, the approach involved directly achieving point-to-point target generation by utilising existing model frameworks and training methodologies. The commonality in these model training methodologies is highlighted by three fundamental steps: constructing datasets, training models based on generative algorithms such as GANs and VAEs, and evaluating the quality of generated outputs. However, point-to-point target generation is primarily outcome-oriented and is particularly suited to liberating designer labour in repetitive design tasks, such as the rapid generation of textile patterns. The outcomes of this approach are relatively uniform and uncontrollable, lacking designer interfaces for convenient adjustments of generated content. With the ongoing integration of AIGC and design, future generative design tools for exploring design space will be further developed and gradually emerge. As Davis et al. (Citation2023), based on creativity-driven divergent and convergent thinking, constructed a clothing design generation system called Fashion, where designers can place inspirational clothing patterns on a two-dimensional panel. The generation system explores the design space along the X and Y axes, producing controlled interpolated images to assist designers in creative ideation. These applications not only incorporate designer control and requirements but also establish continuous feature generation adjustment mechanisms, enabling controlled exploration of the potential design space, effectively facilitating the exploration of design solutions.
自生成模型构建以来，计算智能领域一直在持续优化，以提高生成质量和效率。在设计领域的最初应用中，该方法涉及通过利用现有模型框架和训练方法直接实现点对点目标生成。这些模型训练方法的共同之处在于三个基本步骤：构建数据集，基于生成算法（如 GANs 和 VAEs）训练模型，以及评估生成输出的质量。然而，点对点目标生成主要是以结果为导向的，特别适用于解放设计师在重复设计任务中的劳动，例如快速生成纺织图案。这种方法的结果相对统一且不可控，缺乏设计师界面以便对生成内容进行方便调整。随着人工智能与设计的不断整合，未来用于探索设计空间的生成设计工具将进一步发展并逐渐出现。如戴维斯等人所述。 （2023 年），基于以创造力驱动的分散和收敛思维，构建了一个名为时尚的服装设计生成系统，设计师可以将灵感服装图案放置在二维面板上。该生成系统沿着 X 和 Y 轴探索设计空间，生成受控的插值图像，以帮助设计师进行创意构思。这些应用不仅融入了设计师的控制和需求，还建立了连续的特征生成调整机制，使得能够对潜在的设计空间进行受控探索，有效促进设计解决方案的探索。

6.1.2. From parameterization-based generation to design cognition-based generation
6.1.2. 从基于参数化的生成到基于设计认知的生成

Thoring, Huettemann, and Mueller (Citation2023) explores and indicates that by establishing potential connections between generative artificial intelligence and design knowledge, designers will assume an ‘enhanced’ role with the support of AIGC. This partially answers the symbiotic or replacement question regarding the relationship between AIGC and designers. Leveraging the deep learning capabilities of AIGC, machine cognition can be established through setting and training, such as employing the ‘S (stimulus) - O (object) - R (response)’ framework to correspond to ‘input-black box-output,’ thereby achieving replication of human cognition (Tao, Gao, and Yuan Citation2023).
Thoring, Huettemann 和 Mueller（2023）探讨并指出，通过建立生成人工智能和设计知识之间的潜在联系，设计师将在 AIGC 的支持下承担“增强”角色。这在一定程度上回答了关于 AIGC 与设计师之间关系的共生或替代问题。利用 AIGC 的深度学习能力，可以通过设置和训练建立机器认知，例如采用“S（刺激）-O（对象）-R（响应）”框架对应于“输入-黑盒-输出”，从而实现对人类认知的复制（Tao，Gao 和 Yuan 2023）。

As mentioned earlier, early design generation tools were monolinear, outputting specific generated content based on inputs to the training dataset, which functioned excellently as a generative model. However, as design tools, they were deficient in incorporating theoretical foundations related to design cognition, thinking, processes, and other methodologies. This often resulted in generated outcomes lacking in design insights. Numerous theoretical frameworks have been established in the evolution of the design domain, including design thinking theory which guides the design process, analogical design theory aiding in inspirational ideation, and Kansei engineering contributing to understanding user needs and creating user-centric products. These theories and methodologies help designers in gaining better insights and implementing effective practices. Integrating design thinking with computational thinking represents a crucial aspect of embedding AIGC in the design domain. Some existing research has transitioned from production tools towards roles as design assistants. For instance, Wu (Citation2023), drawing on the Function-Behaviour-Structure (FBS) theory, structured human complex cognitive knowledge of frontal automotive morphology. He developed a cognitive mental model of frontal automotive styling along three dimensions: design intent, drawing behaviour, and functional structure. This approach facilitated the mapping of cognitive design space to machine computational space. With the development and widespread application of AIGC, future research is expected to increasingly integrate generative capabilities with design thinking. This evolution repositions generative models not merely as tools for repetitive tasks but as design applications aligned with cognitive thinking. This advancement further promotes effective collaboration between humans and machines.
正如前面提到的，早期的设计生成工具是单线性的，根据训练数据集的输入输出特定生成的内容，作为生成模型表现出色。然而，作为设计工具，它们在整合与设计认知、思维、流程和其他方法论相关的理论基础方面存在不足。这经常导致生成的结果缺乏设计洞察力。在设计领域的发展过程中建立了许多理论框架，包括指导设计过程的设计思维理论、辅助灵感构思的类比设计理论以及有助于理解用户需求和创建以用户为中心的产品的感性工程。这些理论和方法帮助设计师获得更好的洞察力并实施有效的实践。将设计思维与计算思维相结合代表了将人工智能与设计领域融合的关键方面。一些现有研究已经从生产工具转变为设计助手的角色。 例如，吴（2023）借鉴功能-行为-结构（FBS）理论，对前部汽车形态的人类复杂认知知识进行了结构化。他沿着设计意图、绘图行为和功能结构三个维度发展了前部汽车造型的认知心理模型。这种方法促进了认知设计空间与机器计算空间的映射。随着 AIGC 的发展和广泛应用，未来的研究预计将越来越多地将生成能力与设计思维整合在一起。这种演进将生成模型重新定位为与认知思维一致的设计应用，而不仅仅是重复性任务的工具。这一进步进一步促进了人类与机器之间的有效协作。

6.1.3. From monolithic design generation to a full-process closed-loop design system
6.1.3. 从单片设计生成到全流程闭环设计系统

Design is a multi-step, multi-task process, as outlined by design thinking theory, involving empathising, defining, ideating, prototyping, and testing, with continuous iterations and optimizations between these stages. Confronted with intricate design challenges, the development of comprehensive and integrated design systems is advantageous for organising global and local features and coordinating relationships among various design elements (Charnley, Lemon, and Evans Citation2011). This systematic approach promotes more effective collaboration among designers, ensuring consistency in design decisions and achieving a comprehensive and systematic resolution to complex problems.
设计是一个多步骤、多任务的过程，正如设计思维理论所概述的那样，涉及到共情、定义、构思、原型制作和测试，在这些阶段之间进行持续的迭代和优化。面对复杂的设计挑战，开发全面和整合的设计系统有利于组织全球和本地特征，并协调各种设计元素之间的关系（Charnley，Lemon 和 Evans 2011）。这种系统化的方法促进了设计师之间更有效的协作，确保设计决策的一致性，并实现对复杂问题的全面系统化解决方案。

Research indicates that most AIGC applications in the design domain focus on design generation, i.e. generating specific content according to the target task, such as stylised clothing generation (Kato et al. Citation2018), GUI design generation (Zhao et al. Citation2021), etc. The insights revealed that AIGC also plays a significant role and holds strong potential in design processes such as creative divergence, assistance and advice, evaluation and feedback, which support various design tasks. Therefore, integrating AIGC with the design domain and expanding from basic design generation to the construction of comprehensive design systems is a major emerging trend. As Huang et al. (Citation2023) introduced a design assistance system supporting automotive exterior modification that merges a design generator, design evaluator, and decision optimiser, establishing an all-in-one design system. Lin and Martelaro (Citation2024) developed the AIGC creative integration system ‘Jigsaw,’ enabling designers to intuitively combine various basic models for multi-dimensional design output. Future endeavours should further incorporate the multi-step design processes into design generation systems, constructing modular components for various tasks, thereby promoting design coherence and integrating design resources.
研究表明，在设计领域中，大多数 AIGC 应用集中在设计生成方面，即根据目标任务生成特定内容，例如风格化服装生成（Kato 等，2018 年），GUI 设计生成（Zhao 等，2021 年）等。洞察显示，AIGC 在创意分歧、辅助和建议、评估和反馈等设计过程中也发挥着重要作用，并在支持各种设计任务方面具有强大潜力。因此，将 AIGC 与设计领域整合，并从基本设计生成扩展到构建全面设计系统是一个重要的新兴趋势。正如黄等人（2023 年）介绍的支持汽车外观修改的设计辅助系统，融合了设计生成器、设计评估器和决策优化器，建立了一体化设计系统。Lin 和 Martelaro（2024 年）开发了 AIGC 创意整合系统“拼图”，使设计师能够直观地组合各种基本模型，实现多维设计输出。 未来的努力应进一步将多步设计过程纳入设计生成系统，为各种任务构建模块化组件，从而促进设计的连贯性并整合设计资源。

6.2.1. Explore application mechanisms for integrating AIGC with the design
6.2.1. 探索将 AIGC 与设计集成的应用机制

Combining AIGC with the design domain is expected to reshape the new design process. Generative AI facilitates automating tedious, repetitive, and inefficient tasks in design, allowing designers to focus more on creativity and decision-making, accelerating the design process and enhancing efficiency (Liu, Fu, and Li Citation2023). However, the framework mechanisms for the fusion of these two domains remain in the exploratory stage. AIGC demonstrates excellent potential in all aspects of the design processes: creative divergence, design generation, assistance and advice, evaluation and feedback. It is mostly based on traditional design processes and has developed generative tools for specific aspects and tasks of design. In the future, AIGC is expected to transcend the traditional framework of design and pioneer a new method of design. Deep learning-based design evaluation prediction tools are typical. For instance, visual modelling and saliency analysis are employed to predict user click probabilities on UIs in interface design (Schoop et al. Citation2022). This approach breaks through the lag of traditional design evaluation and provides design insights based on existing data. Further combined with design generation, it facilitates automatic optimisation and fast iteration of content (Heyrani Nobari, Fathy Rashad, and Ahmed Citation2021). Conversely, unlike the construction of end-to-end design generation tools, general-domain oriented AIGC tools such as ChatGPT and MidJourney incorporate extensive database resources and exhibit strong generative capabilities. Developing application frameworks through open interfaces allows them to move to the design domain. For example, MidJourney is guided to design generation through four steps: initial prompt, prompt adjustment, style refinement, and variant selection (Cheng Citation2023). As well as constructing cue word frameworks such as ‘Few-Shot’ (Ma et al. Citation2023) and task decomposition (Wang et al. Citation2023) in the Large Language Model (LLM) to assist designers in creative diffusion. Therefore, numerous possibilities exist for future exploration of the mechanisms of combining AIGC and design. On one hand, we can consolidate the existing application mechanism of integrating AIGC with the traditional design process, and develop more effective tools for design field and task refinement. On the other hand, we can try to explore the new mechanism and conceptualise the update and development of the design process led by AIGC.
将 AIGC 与设计领域相结合预计将重塑新的设计流程。生成式人工智能有助于自动化设计中繁琐、重复和低效的任务，使设计师能够更多地专注于创造力和决策，加快设计过程并提高效率（刘、付和李，2023 年）。然而，这两个领域融合的框架机制仍处于探索阶段。AIGC 在设计过程的各个方面展现出极大的潜力：创意分歧、设计生成、辅助和建议、评估和反馈。它主要基于传统设计流程，并为设计的特定方面和任务开发了生成工具。未来，预计 AIGC 将超越传统设计框架，开创一种新的设计方法。基于深度学习的设计评估预测工具是典型的。例如，在界面设计中，利用视觉建模和显著性分析来预测用户对用户界面的点击概率（Schoop 等人，2022 年）。这种方法突破了传统设计评估的滞后，并基于现有数据提供设计见解。 进一步结合设计生成，它促进了内容的自动优化和快速迭代（Heyrani Nobari，Fathy Rashad 和 Ahmed 2021）。相反，与端到端设计生成工具的构建不同，诸如 ChatGPT 和 MidJourney 之类的通用领域导向的 AIGC 工具融入了广泛的数据库资源，并展现出强大的生成能力。通过开放接口开发应用框架使它们能够进入设计领域。例如，MidJourney 通过四个步骤引导设计生成：初始提示，提示调整，风格完善和变体选择（Cheng 2023）。以及在大型语言模型（LLM）中构建诸如“Few-Shot”（Ma 等人，2023）和任务分解（Wang 等人，2023）的提示词框架，以协助设计师进行创意扩散。因此，未来探索将 AIGC 与设计相结合的机制存在许多可能性。一方面，我们可以巩固将 AIGC 与传统设计流程整合的现有应用机制，并开发更有效的设计领域和任务细化工具。 另一方面，我们可以尝试探索由 AIGC 主导的设计过程的更新和发展的新机制并加以概念化。

6.2.2. Incorporating pluralistic methodologies
6.2.2. 融合多元方法论

Design is a multifaceted research domain, enriched by numerous mature theories and methodologies developed to scientifically guide the design process across various tasks and activities. The integration of design and AIGC signifies the convergence of design thinking and computational thinking, necessitating the incorporation of diverse design methodologies to guide the development of AIGC applications. However, based on the review's findings, most existing research on integrating AIGC with design primarily focuses on training generative models to fulfil design tasks. While such studies contribute to alleviating labour-intensive functions in design, they often fall short in addressing creative tasks, such as exploration in design space. A minority of research literature has successfully mapped from human cognitive to computational generative space by explicitly integrating established design theories as the foundation for application development. In future research, it is essential to categorise generative tasks, systematically summarise and map corresponding design theories, and employ a diverse theoretical foundation to guide the construction of generative applications.
设计是一个多方面的研究领域，通过许多成熟的理论和方法论的丰富，科学地引导设计过程跨越各种任务和活动。设计与 AIGC 的整合标志着设计思维和计算思维的融合，需要将多样化的设计方法论纳入，以指导 AIGC 应用的开发。然而，根据审查结果，大多数现有关于将 AIGC 与设计整合的研究主要集中在训练生成模型以完成设计任务上。虽然这些研究有助于减轻设计中的劳动密集型功能，但它们往往在解决创造性任务方面表现不佳，比如在设计空间中的探索。少数研究文献成功地将人类认知映射到计算生成空间，明确将建立的设计理论作为应用开发的基础。在未来的研究中，有必要对生成任务进行分类，系统总结和映射相应的设计理论，并采用多样化的理论基础来指导生成应用的构建。

6.2.3. Obstacles and negative impacts
6.2.3. 障碍和负面影响

Perceived anxiety and risk negatively impact the willingness to use AIGC design tools (Li Citation2024). Generative variability is an essential feature of AIGC, i.e. based on the same input, the output content has diversity and variability, which, although it violates the common principle of Human–Computer Interaction (HCI) that requires a system to respond consistently to the user's inputs (Weisz et al. Citation2024), it is also to a certain extent compatible with design as a structurally ill-posed problem, which realises the exploration of the potential space of design through generative diversity. However, the prerequisite is that systems and designers need to understand the probabilistic nature and generative capabilities of AIGC to realise effective interaction between the two. The HCI community proposes a principle for the construction of AI applications: human-centered AI (HAI) states that intelligent systems need to be constructed with humans as their stakeholders in mind and that it is, therefore, necessary to build models of machine behaviour that correspond to models of the human mind, avoiding the ‘black box’ state (Riedl Citation2019). In addition, Explainable Artificial Intelligence (XAI) states that building AI collaborative partners that human users can understand, trust, and manage effectively ensures fairness in decision-making and improves system robustness (Barredo Arrieta et al. Citation2020; Samek and Müller Citation2019). It is necessary to build design generation applications based on HAI and XAI considerations in the future to effectively overcome existing barriers and improve the transparency and interpretability of applications.
感知焦虑和风险会对使用 AIGC 设计工具的意愿产生负面影响（Li 2024）。生成变异性是 AIGC 的一个重要特征，即基于相同的输入，输出内容具有多样性和变异性，尽管这违反了要求系统对用户输入保持一致响应的人机交互（HCI）的共同原则（Weisz 等人，2024 年），但在一定程度上与设计作为一个结构上不明确的问题相容，通过生成多样性实现对设计潜在空间的探索。然而，前提是系统和设计师需要了解 AIGC 的概率性质和生成能力，以实现两者之间的有效交互。人机交互社区提出了构建人工智能应用的原则：以人为中心的人工智能（HAI）指出智能系统需要以人类为利益相关者构建，并且因此有必要建立与人类思维模型相对应的机器行为模型，避免“黑匣子”状态（Riedl 2019）。 此外，可解释人工智能（XAI）指出，构建人类用户能够理解、信任和有效管理的 AI 协作伙伴，可以确保决策的公平性并提高系统的稳健性（Barredo Arrieta 等，2020 年；Samek 和 Müller，2019 年）。未来有必要基于 HAI 和 XAI 的考虑构建设计生成应用程序，以有效克服现有障碍并提高应用程序的透明度和可解释性。

AIGC relies on generative models, which are processed based on training pre-existing datasets. There is a correlation between the content output and the training input, which leads to ethical and copyright issues for AIGC (Shi et al. Citation2023). First, generative AI may be misled by datasets containing erroneous and biased content and thus stereotypes in the output, such as sexism, racial discrimination, geographical discrimination, etc (Weidinger et al. Citation2021). This problem can be overcome through input screening, training optimisation, and output control. Develop screening criteria for datasets to create excellent and ethical datasets to reduce biased content. Optimise the algorithmic model to incorporate a control component to censor and block sensitive content. Build an evaluation mechanism for the output to detect and control the output content and to adjust the generative system promptly.
AIGC 依赖生成模型，这些模型是基于训练预先存在的数据集进行处理的。内容输出与训练输入之间存在相关性，这导致了 AIGC 的道德和版权问题（Shi 等，2023 年）。首先，生成式人工智能可能会被包含错误和偏见内容的数据集误导，从而在输出中出现刻板印象，如性别歧视、种族歧视、地理歧视等（Weidinger 等，2021 年）。这个问题可以通过输入筛选、训练优化和输出控制来解决。制定数据集的筛选标准，创建优秀和道德的数据集以减少偏见内容。优化算法模型，加入控制组件以审查和阻止敏感内容。建立一个评估机制，用于检测和控制输出内容，并及时调整生成系统。

In addition, the migration and reconstruction relied on by generative models may lead to the output of similar content not controlled by designers, which involves copyright and plagiarism issues, e.g. the migration of styles of branded clothing, the learning of hand-drawn sketches by designers, etc., which may lead to illegal plagiarism and impersonation through the misuse of generative models. To address this problem, the corresponding visual and semantic similarity evaluation system needs to be further improved and the detection method needs to be constructed in the future.
此外，生成模型依赖的迁移和重建可能导致设计师无法控制的类似内容的输出，涉及版权和抄袭问题，例如品牌服装风格的迁移，设计师手绘草图的学习等，这可能通过滥用生成模型导致非法抄袭和冒名顶替。为解决这一问题，未来需要进一步完善相应的视觉和语义相似性评估系统，并构建检测方法。