With the rise of Generative Adversarial Networks (GANs), AI has increasingly become a partner to human designers in co-creating cultural artifacts. While generative models have been applied in various creative tasks across disciplines, a theoretical foundation for understanding human-GAN collaboration is yet to be developed. Drawing from the mixed-initiative co-creation community, we propose a preliminary framework to analyze co-creative GAN applications. We identify four primary interaction patterns: Curating, Exploring, Evolving, and Conditioning. The suggested framework enables us to discuss the afordances and limitations of the diferent kind of interactions underlying co-creative GAN applications.
co-creation with AI agents [8]. For example, Spoto and
Oleynik [13] propose a framework to map the primary
With the recent development in human-AI collaborative interactions taken in the co-creative process between
applications, deep generative models have become poten- humans and computational agents [13]. Muller et al. [14]
tial co-designers to humans in creative tasks. Especially further expand the above-mentioned framework by
inGenerative Adversarial Networks (GANs) have gained at- cluding actions specific to generative AI interfaces. While
tention from designers and non-programmers across cre- Muller et al.’s expanded framework provides a tool for
ative disciplines such as architecture, fashion, computer analyzing generative AI interfaces, its extensive action
games, and art [
While technological advancements underlying GANs GANs’ interactive capabilities accurately.
progress rapidly [
Research in mixed-initiative co-creation (MI-CC) [11, models [14]. Based on the MI-CC frameworks, we sug12] ofers a promising starting point to bridge this knowl- gest a preliminary framework for in-depth analysis of edge gap. The research area of MI-CC focuses on how hu- co-creative GAN applications. We show that the sugmans and machines can co-create together, including the gested taxonomy lets us analyze emerging patterns in the interaction with GANs. We limit our study to GANs because the majority of examples in the field of generative design use them, but we believe that our framework applies to other deep generative models. More specifically, the proposed set of actions supports the mapping Joint Proceedings of the ACM IUI Workshops 2022, March 2022, Helsinki, Finland nEvelop-O imgr@itu.dk (I. Grabe); migd@itu.dk (M. González-Duque); sebr@itu.dk (S. Risi); jicz@itu.dk (J. Zhu)
© 2022 Copyright © 2022 for this paper by its authors. Use permitted under CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g CCreEatUiveRCoWmmoornksLsichenospeAPttrribouctioene4d.0iInntgersna(tiConEal U(CCRB-YW4.0S)..org) of interactions with generators via latent codes. Hence, proposing new GAN applications [21, 22, 23, 24] or sugthese actions are applicable to other generative models, gesting algorithmic methods to control GANs without such as variational autoencoders (VAEs), consisting of a pre-defined use cases [25, 26, 27]. As part of a study on generator network with outputs that are sampled based AI-augmented typeface design, Zeng et al. [28] frame a on alterable latent codes. GAN as a source of inspiration in a creativity model. To
By synthesizing existing GAN applications and MI- date, the only (systematic) review in the area is Hughes
CC theory, our work aims to shed light on how GANs et al.’s [
In summary, this paper presents our preliminary framework for mapping the interaction in co-creative GAN The research area MI-CC investigates interactive proapplications and discusses the co-creative afordance of cesses in which both human and machine contribute the emerging interaction patterns. The rest of the paper proactively to producing artifacts [12]. MI-CC appliis structured as follows. In the next section, we sum- cations can be placed on a continuum, describing the marize related work. Then, we present the framework extent to which human and computational agent take supported by exemplary co-creative GAN applications initiative in the creative process [29, 30]. To analyze the drawing on existing literature. We discuss how the iden- initiatives taken in such collaboration, several theoretical tified interaction patterns support co-creativity, followed approaches have been suggested. Spoto and Oleynik [13] by a conclusion and outlook on future work. proposed to decompose the MI-CC process into seven primary actions: ideate, constrain, produce, suggest, select, assess, adapt. The actions can be used to map the 2. Related Work “creative flow” [13] between human and computer as a graph. By mapping the series of actions taken by the two In this section, we summarize theoretical work on under- agents, the authors outline the creative (iterative) process standing co-creative GAN applications and approaches for more than 70 works ranging from game level creation to map MI-CC interactions. to manufacturing design systems. When it comes to including AI agents in co-creation with human designers, 2.1. Understanding human-AI new interaction possibilities emerge, requiring comprecollaboration with GANs hensive analysis frameworks. Muller et al. [14] adapt the notation of Spoto and Oleynik’s framework to processes including generative AI agents. They add four actions that are specific to generative models, such as learn to describe the computer’s training process. Their extended framework can be used to visualize “human-AI interaction patterns in the generative space” [14, p.1].
Our work aims to tailor these MI-CC methods to analyze co-creative GAN applications. We build on existing mapping methods suggested by MI-CC to outline common interaction patterns in co-creation between human and GAN.
At the intersection of HCI and AI research, designing
human-AI interaction remains an open issue due to the
complexity of AI models and their lack of
interpretability [9, 15]. This includes designing co-creative GAN
applications. GANs learn to generate content from latent
features, but this mapping is not explainable and
blackbox, making it dificult to control the desired output after
training [
methods to understand how GANs function in cocreation. As a starting point of this study, we reviewed Initialize. To initialize refers to setting up the GAN, GAN applications in salient design domains. Reviewed such as choosing a dataset and a model architecture. studies ranged from prompting GANs to invent new art Learn. To learn describes the process of internalizing styles [31], over breeding game levels [21], to adjusting information. In line with Muller et al. [14], this refers fashion items as part of a person’s outfit [25]. To make to the training GANs on a chosen data distribution. For sense of the surveyed GAN studies, we leveraged exist- the human counterpart, the action would correspond to ing MI-CC frameworks to map the actions in co-creative learning new skills or adapting to a new domain. tasks. They allowed us to analyze how GANs get involved Constrain. To constrain refers to the process of speciin creative processes. In an iterative manner, we analyzed fying desired characteristics of the target artifacts, hence studies and revised the frameworks to derive one that is restricting the conceptual space. applicable across co-creative GAN applications. During Create. To create describes the action of generating this process, we focused on how the frameworks’ actions new (candidates of) artifacts. The number of artifacts to reflect the GANs’ technical functioning. Along the way, be created can vary from one to many. predominant flows in how humans and GANs interact Select. To select refers to the action of choosing one or in co-creation stood out among the studies, which we more artifacts and excluding others from the further prorefer to as interaction patterns. This section presents the cess. The action encompasses diferent selection mechasuggested framework and demonstrates the four primary nisms including the structuring of subsets such as rankinteraction patterns we identified with it. ing.
Adapt. To adapt describes the act of making changes 3.1. Framework to existing artifacts. The agents can edit artifacts directly or adjust their representation, namely the latent code, Our framework builds on the framework suggested by that is used to create a new artifact. Spoto and Oleynik [13], further extended to map co- Combine. To combine describes the process of concreation with generative models by Muller et al. [14]. structing a new artifact that inherits parts of existing
We adapt their proposed actions to a minimal set re- artifacts. As with adapting, this is done via the artifacts’ quired to describe co-creative GAN applications. For latent code. example, Muller et al. distinguish between producing one The actions described above present a minimal set sufiartifact and suggesting a set of artifacts to choose from cient to describe the core interactions of co-creative GAN as separate actions. We combine them into one category applications. For example, we left out the action ideate, called create, as the artifact sampling from the GAN de- which describes creating high-level concepts in existing picts the same action in either case. In addition, we add frameworks [13, 14]. As we find that conceptual ideation new actions, such as initialize, which we find can present is often expressed by conditions imposed on a GAN, the a crucial design choice taken by human agents [24]. The action constrain covers it in our framework, aligned with ifnal set of actions illustrates the co-creation as a graph the GANs’ algorithmic function. The derived action set mapping along the two axes of agents and actions. This aims to align with the algorithmic properties of GANs section defines the agents making up the horizontal axis as well as the interactive abilities of the human designer. and explains what the actions on the vertical axis stand Mapping interactions with the framework may help nonfor. technical users better understand a co-creative partner’s algorithmic actions and show how the human designer’s 3.1.1. Agents actions embed into it. The set of core actions allows us to map the interactions of the co-creative GAN applications we surveyed.
Using the above framework of agents and actions, we identified four interaction patterns which we present along with an exemplary selection of co-creative GAN applications in which they appear. We reviewed studies that propose models for co-creation with GANs tested in user studies, such as a GAN that allows game designers to generate Mario levels [21]. Besides that, we also considered approaches that are not explicitly embedded into use cases yet but suggest applicable models that The agents conducting the actions are (1) a human designer and (2) a computational system including a GAN as the main technology, which might be supported by other algorithms facilitating the interaction with it. 3.1.2. Actions The actions are to be understood as activities that influence the artifact or the other agent directly. How actions can be executed depends on several factors, such as the design of the interface, the implementation details of a model, or the abilities of a human agent. Seven actions are used to map the interaction between the two agents, as described below. could support co-creative GAN applications. For exam- They replace the production step of design drawing in a ple, Yildirim et al. [26] trained a GAN to generate dress fashion design workflow with GAN creations. Then, the designs based on given features such as color. Among all output design is given to pattern makers. studies, we found four underlying interaction patterns Besides selecting the training data, the model architecfor which we show the graphical notations in Figure 1. ture can also pose a design factor. Elgammal et al. [31] Interaction patterns are not mutually exclusive and can demonstrate that with art-generating Creative Adversarbe combined. Table 1 provides an overview of how the ial Networks (CANs). They apply a style ambiguity loss surveyed examples apply the patterns. when training the networks on an art dataset assigned to diferent art styles as categories. As a result, the CAN 3.2.1. Curating learns to output artifacts that deviate from existing art styles. While there are no limitations to the complexThe most straight-forward interaction pattern is catego- ity of designing the initial GAN set-up, Curating is the rized as Curating. After being initialized by a human simplest form for GANs to partake in the subsequent designer, a GAN learns a to generate artifacts by being creative process. trained through gradient descent. Then, a set of outputs is created by sampling from the GAN, from which the 3.2.2. Exploring human agent selects a set of final artifacts. Note that the selection can also include all generated artifacts. After The interaction pattern Exploring describes an interacthe training, the GAN is the only one influencing the tion where the human iteratively adapts artifacts that creation process. It does not receive any input from the are created by sampling from the GAN. This is typically human side. The main creative human input is given done by introducing alterations via the latent representaduring the initialization. Hence, choosing a dataset and a tions of artifacts. Altering the latent code of an artifact model plays a crucial role in the human agent’s initiative. can be understood as providing a recipe to create a new
Curating has been applied to produce new artifacts by artifact. By steering the following creation process, the adding variation to an existing dataset in several domains. designer guides the exploration path through the space For instance, fashion designers might purposely choose a of possible designs. The latent space can be explored by collection of clothing designs as training distribution.1 In traversing along diferent directions. These directions that way, they determine the overall style of newly gen- cause diferent changes to the output features. Users can erated artifacts to which they aim to add novelty through control the change in direction by e.g. moving a slider. the GAN’s variation. Kato et al. [32] train a GAN to pro- We present three common directions to adapt artifacts duce outputs of similar quality to human-made designs. below. Figure 2 provides a visual example per direction.
Interpolation. One direction to move an artifact
through latent space is towards other artifacts in the
1https://www.vogue.com/fashion-shows/fall-2020menswear/acne-studios
space. In doing so, one explores intermediate alterna- gled during the training process. In practice, diferent
tives between existing artifacts. Among other methods, attributes link to diferent layers and sliders let users
Schrum et al. [21] apply interpolation to explore the de- manipulate target features. On top of that, Artbreeder
sign space of GANs trained to create tile-based rooms allows users to introduce semantic directions based on
for the game The Legend of Zelda. By traversing be- sample images containing target features.
tween two diferent level designs, the game designer can The features along which users explore the conceptual
prompt the GAN to create new designs lying in between. space can also be combined. For example, CREA.blender
CREA.blender [33], a tool that tests users’ creativity in SDG [
Latent directions. Instead of altering GAN-generated the interaction, two sliders below each image let the huartifacts in relation to one another, they can also be man agent control how much of its content and style feed adapted by directly increasing or decreasing their latent into the resulting design. Hence, the simple interpolation variables. As a result, the artifact moves along the GAN’s of artifacts expands to consider semantic properties. latent directions. In another part of Schrum et al.’s study, Bau et al. [38] also bridge the gap between the interusers adjust a game level design by moving a slider per nal representation of GANs and interpretable concepts. latent variable. Because the latent variables do not cor- By analyzing how the latter encode in the former, they respond to visual output features, changes to the output propose GANDissect, a method to identify latent units design are not necessarily explainable, which users might in the generator’s layers that activate certain aspects in experience as frustrating when having a target change the output image. Taken to practice in the application in mind [21]. GANPaint [39], the authors propose an interface that
Semantic features. The adaption along directions in lets users add, delete, or alter aspects in images. Using a latent space that represent semantic attributes is better brush to indicate target locations, or pixels, in the image, understandable. After finding such an attribute vector the human agent can adapt GAN-generated artifacts by in the latent space, an artifact’s latent code can traverse indirectly de-/activating the corresponding latent units. along it. That influences the appearance of the attribute By representing actual photographs as a GAN’s generain the artifact. Research in the domain of semantic face tion, the changes can be applied to existing images. This editing [34, 35] has widely explored this technique to allows a human agent “to manipulate a photograph not alter facial features. For example, the editing function with physical colors, but with abstract concepts such as in the interface of the Composite Generating GAN (CG- object types and visual attributes” [39, p.2]. GAN) [23] supports interaction in that manner. Human agents adapt generated faces by adding or subtracting 3.2.3. Evolving certain facial traits. They may choose to lock other facial features during that process to avoid correlated changes. In the third identified pattern, Evolving, the human se
ArtBreeder [36], a website where users can make GAN- lects artifacts for the computational agent to adapt and
generated art, applies another approach to disentangle combine in the following step. Within the field of
evolusemantic features, resulting in a similar control to CG- tionary computation, these actions are also referred to as
GAN for a faces generator. By applying the BigGAN [37] mutation and recombination, respectively. This
interacor StyleGAN [
Bontrager et al. [41] let human agents evolve faces, functions. During training, parts of the latent code are shoes, and chairs in that manner. Users select artifacts assigned to the three features. With this model, a user from the first creation step. Their latent codes are then could pre-define the shape of a dress in a hand drawing. recombined though crossover and mutated by adding The drawing is then captured as a 512-dimensional mask Gaussian noise. With the adapted latent codes, new items and included into the input given to the GAN. Similarly, are sampled and the process repeats. With the power Xin and Arakawa [43] apply a GAN model that takes to guide the process through selecting artifacts to sur- a contour image as input as part of their object design vive, the human agent optimizes the resulting artifacts system. In that, the human designer can determine an according to their preferences. It is the computational outline before the GAN generates the artifact, such as the agent that adjusts the artifacts primarily. However, the contour of a shoe. Zhao and Ma [44] note that providing human agent can influence that, too, by specifying the a conditional GAN with a hand-drawn constraint can rules for this traversal search [12]. In practice, this would cause dificulties when the drawing style difers from the imply setting algorithmic parameters such as mutation distribution of training drawings [44]. or crossover rate. Schrum et al. demonstrated that in the Setting landmarks of a drawing may help further define evolving functionality of their study [21]. Here, users a target design. Therefore, the authors apply a second select generated game levels to be recombined and mu- generative network in the constraint step, that transforms tated for the next generation, also referred to as selective those landmarks into a compensation image. Together breeding. By moving a slider, the user sets the rate for with the user’s drawing, it acts as a reference to guide the applied polynomial mutation. In the evolving process the generation by the main GAN. The example shows underlying CG-GAN, the user can choose along which that constraints can consist of multiple inputs, as is also semantic features the algorithms should mutate artifacts. the case in the following case for clothing design sysSimilar directions for traversing the latent space as pre- tem suggested by Zhu et al. [25]. The authors let human sented in section 3.2.2 could also be specified for the designers guide the fashion design flow by giving an computational agent. The human agent might even se- instruction in text form. First, a segmentation mask is lect them during the design process. However, as the derived from the photo of the person to be dressed. Tohuman agent does not directly adapt artifacts in this pat- gether with the text encoding of an outfit description, tern, the directions do not concern the interaction and the mask is given to their GAN model as input. This are not stated here. allows them to dress a person according to a human’s idea. The GAN contributes creative initiative through 3.2.4. Conditioning unpredictable variation.
man constrains the desired artifact characteristics before outputs are created through sampling from the GAN.
to be understood as an elementary set of interactions between human and GAN. As some of the reviewed applications show, they can be combined to create more advanced co-creative processes.
For instance, Xin and Arakawa [43] combine Evolving and Conditioning in their suggestion of an object design system. Figure 4 shows the interaction pattern for the system trained on datasets of shoes, handbags, and other fashion items. In the first step of their user study, users evolved shoe designs in an IGA without the option to constrain the design space, following simply the Evolving pattern. Secondly, they extend the generation process so that users could prompt the GAN’s creation with a contour image of the desired shoe (see Figure 3). By adding Conditioning to the Evolving interaction pattern, users can better derive designs that resemble the target sneaker, as the example in Figure 4 shows.
Instead of allowing the user to pre-limit the conceptual space of a GAN, Zaltron et al. [23] give users the chance to adapt generated images. The interaction underlying their proposed CG-GAN combines the Evolving and Exploring pattern, as displayed in Figure 5. First, the model lets users evolve faces similar to Xin and Arakawa’s IGA for object design. But during the process, users can edit artifacts by stepping out of the evolutionary loop. In doing so, users fine-tune a generated design by traversing along semantic facial features with the help of sliders. The evolved artifact is then added back into the loop.
Oppositely, we can observe the Exploring pattern being expanded with Evolving functionalities in CREA.blender [33]. In its simplest form, CREA.blender asks users to develop animal-like shapes based on preselected basis images. Then, in the open-play mode, users can themselves select basis images by replacing the current ones. With the selected images as a starting point, they explore new images. In comparison to the “more goal-oriented creative tasks” [33, p.4], letting the user take part in the selection supports their “open-ended creativity” [33, p.4].
isolated process. However, in the case of Conditioning, a user might map the constraint space and learn In this section we discuss the co-creative afordance of its constraint language by iterating the pattern repeatthe interaction patterns and reflect on the potential usage edly. While our preliminary framework does not curof the framework. rently capture this step, this points to the asymmetry
The interaction patterns we identified based on the in our framework, focusing on the GAN-specific techframework allow us to analyze how co-creative GAN ap- nical grounding of actions rather than human-specific plications support co-creativity to diferent degrees. For aspects. example, where Muller et al. [14] identify one interaction Opposed to the one-shot patterns, Exploring and Evolvpattern, we distinguish between Curating and Condition- ing imply iterative interaction between human and GAN. ing. More specifically, they categorize Elgammal et al.’s While human initiative appears only after the first creCAN and Zhao and Ma’s design method into the same in- ation stage in both patterns, they support the co-creative teraction pattern characterized by the population of a so- exchange diferently. Through selection in an Evolving lution space with a generative model. However, the two lfow, the human agent narrows the conceptual space of examples are fundamentally diferent in how they sup- the GAN. Hence, they restrict its creative potential to deport co-creativity, and our proposed adaption allows us veloping a chosen population of designs. In comparison, to make a fine-grained distinction between the two ways Exploring leaves the conceptual space as is but lets the of interacting. Curating, as in the case of CAN, lets the human designer traverse through it. However, the direchuman pre-determine the GAN’s design space through tions taken restrict the exploration. Freely altering the its initialization as an active design choice. Hence, the latent variables allows for a broader exploration than the human agent’s primary input limits the GAN’s creative interpolation between two artifacts, whereas traversing potential before training. Conditioning allows the hu- along semantic features limits the search to given paths. man to narrow the existing design space of the GAN after Overall, Exploring afords co-creativity by simply allowtraining. We categorize Zhao and Ma’s example as such ing the human agent to move in the conceptual space. because the human agent steers the GANs by providing In contrast, Evolving supports a more targeted search constraints, in this case, a drawing and landmarks, after strategy for the user but restricts the GAN’s creative the initial training. While Curating leaves the human potential. agent out of the actual (co-)creative process and gives the As the identified patterns support diferent purposes, GAN complete creative control, Conditioning restricts their mapping allows for a reflection about their applithe GAN’s creativity by assigning the human agent more cation and combination. When designing a process that creative authority along the way, hence allowing for more starts from pre-defined characteristics, modeling the inco-creativity. By committing to a human-set frame, the teraction according to the Conditioning pattern would GAN loses parts of its creative potential, leading to a let humans set the scope right from the beginning of a trade-of between novelty and typicality with regards to creation process. But when designing for a scenario that humans’ expectations. requires targeted search for distinguished properties in
We regard Curating and Conditioning as one-shot ap- a design space, Evolving might be the pattern of choice. proaches since the computational agent does not recall Exploring, on the other hand, could support a more inspipreviously created artifacts, making each iteration an rational excursion through artifact possibilities. When
Our preliminary framework suggests mapping interactions as a basis for understanding co-creative GAN applications. We identified four primary interaction patterns, which we demonstrated along with examples of existing approaches. The patterns let us understand how diferent GAN applications support co-creativity. This insight can inform how we design co-creative GAN applications.
The preliminary framework contributes to bridging the knowledge gap between machine learning engineers and [26] G. Yildirim, N. Jetchev, R. Vollgraf, U. Bergmann, //ieeexplore.ieee.org/document/9157070/. doi:10.
Generating High-Resolution Fashion Model Images 1109/CVPR42600.2020.00926.
Wearing Custom Outfits, 2019. URL: http://arxiv. [36] J. Simon, Artbreeder, 2022. URL: https://www. org/abs/1908.08847. artbreeder.com/browse. [27] B. Roziere, F. Teytaud, V. Hosu, H. Lin, J. Rapin, [37] A. Brock, J. Donahue, K. Simonyan, Large Scale M. Zameshina, O. Teytaud, EvolGAN: Evolutionary GAN Training for High Fidelity Natural Image Generative Adversarial Networks, 2020. URL: http: Synthesis, arXiv:1809.11096 [cs, stat] (2019). URL: //arxiv.org/abs/2009.13311. http://arxiv.org/abs/1809.11096, arXiv: 1809.11096. [28] Z. Zeng, X. Sun, X. Liao, Artificial Intelligence [38] D. Bau, J.-Y. Zhu, H. Strobelt, B. Zhou, J. B. TenenAugments Design Creativity: A Typeface Fam- baum, W. T. Freeman, A. Torralba, GAN Dissecily Design Experiment | SpringerLink, in: De- tion: Visualizing and Understanding Generative sign, User Experience, and Usability. User Expe- Adversarial Networks, in: Proceedings of the Interrience in Advanced Technological Environments, national Conference on Learning Representations 2019. URL: https://link.springer.com/chapter/10. (ICLR), 2019, p. 19.
1007/978-3-030-23541-3_29. [39] D. Bau, H. Strobelt, W. Peebles, J. Wulf, B. Zhou, [29] S. Deterding, J. Hook, R. Fiebrink, M. Gillies, J. Gow, J.-Y. Zhu, A. Torralba, Semantic photo manipulaM. Akten, G. Smith, A. Liapis, K. Compton, Mixed- tion with a generative image prior, ACM TransInitiative Creative Interfaces, in: Proceedings of actions on Graphics 38 (2019) 1–11. URL: https: the 2017 CHI Conference Extended Abstracts on //dl.acm.org/doi/10.1145/3306346.3323023. doi:10. Human Factors in Computing Systems, ACM, Den- 1145/3306346.3323023. ver Colorado USA, 2017, pp. 628–635. URL: https: [40] H. Takagi, Interactive evolutionary computation: //dl.acm.org/doi/10.1145/3027063.3027072. doi:10. fusion of the capabilities of EC optimization and 1145/3027063.3027072. human evaluation, Proceedings of the IEEE 89 [30] A. Liapis, G. N. Yannakakis, Boosting Computa- (2001) 1275–1296. URL: http://ieeexplore.ieee.org/ tional Creativity with Human Interaction in Mixed- document/949485/. doi:10.1109/5.949485. Initiative Co-Creation Tasks, in: Proceedings of the [41] P. Bontrager, W. Lin, J. Togelius, S. Risi, Deep InterICCC workshop on Computational Creativity and active Evolution, in: International Conference on Games, Paris, 2016. Computational Intelligence in Music, Sound, Art [31] A. Elgammal, B. Liu, M. Elhoseiny, M. Mazzone, and Design : EvoMUSART 2018, 2018.
CAN: Creative Adversarial Networks, Generating [42] G. Yildirim, C. Seward, U. Bergmann, Disentangling ”Art” by Learning About Styles and Deviating from Multiple Conditional Inputs in GANs, in: KDD 2018 Style Norms (2017). URL: http://arxiv.org/abs/1706. Conference AI for Fashion Workshop, 2018. URL: 07068. http://arxiv.org/abs/1806.07819, arXiv: 1806.07819. [32] N. Kato, H. Osone, K. Oomori, C. W. Ooi, Y. Ochiai, [43] C. Xin, K. Arakawa, Object Design System GANs-based Clothes Design: Pattern Maker Is All by Interactive Evolutionary Computation Using You Need to Design Clothing, in: Proceedings of the GAN with Contour Images, in: Human Cen10th Augmented Human International Conference tred Intelligent Systems - Proceedings of KES2019, ACM, Reims France, 2019, pp. 1–7. URL: https: HCIS 2021 Conference, volume 244, Springer Singa//dl.acm.org/doi/10.1145/3311823.3311863. doi:10. pore, Singapore, 2021, pp. 66–75. URL: https://link. 1145/3311823.3311863. springer.com/10.1007/978-981-16-3264-8_7. doi:10. [33] M. Gajdacz, J. Rafner, S. Langsford, A. Hjorth, 1007/978- 981- 16- 3264- 8_7, series Title: Smart C. Bergenholtz, M. M. Biskjaer, L. Noy, S. Risi, Innovation, Systems and Technologies. J. Sherson, CREA.blender: a GAN based casual [44] Z. Zhao, X. Ma, A Compensation Method of Twocreator for creativity assessment, in: Proceedings Stage Image Generation for Human-AI Collaboof the International Conference on Computational rated In-Situ Fashion Design in Augmented ReCreativity (ICCC ’21), 2021, p. 5. ality Environment, in: 2018 IEEE International [34] E. Denton, B. Hutchinson, M. Mitchell, T. Gebru, Conference on Artificial Intelligence and Virtual A. Zaldivar, Image Counterfactual Sensitivity Anal- Reality (AIVR), IEEE, Taichung, Taiwan, 2018, pp. ysis for Detecting Unintended Bias (2019). URL: 76–83. URL: https://ieeexplore.ieee.org/document/ http://arxiv.org/abs/1906.06439. 8613637/. doi:10.1109/AIVR.2018.00018. [35] Y. Shen, J. Gu, X. Tang, B. Zhou, Interpreting the [45] D. Bau, S. Liu, T. Wang, J.-Y. Zhu, A. Torralba, Latent Space of GANs for Semantic Face Editing, Rewriting a Deep Generative Model, in: Proceedin: 2020 IEEE/CVF Conference on Computer Vi- ings of the European Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, Seat- sion (ECCV), 2020. tle, WA, USA, 2020, pp. 9240–9249. URL: https: