The Prompt Artists

AI tools have played a prominent role in creativity support tool (CST) research (Frich et al., 2019; Hwang, 2022). AI-based CST systems have been produced to support artistic generation in domains such as fashion and product design (Sbai et al., 2018; Jeon et al., 2021; Quanz et al., 2020), music creation (McCormack et al., 2019; Louie et al., 2020; Huang et al., 2020), drawing (Davis et al., 2015, 2016; Oh et al., 2018; Karimi et al., 2019), visual design and story-boarding (Zhao et al., 2020; Shi et al., 2020), and storytelling (Hodhod and Magerko, 2016; Perrone and Edwards, 2019). Most of these tools support the artistic implementation process as either production aids (i.e., tools that perform most of the work of generating the art, e.g., generative text-to-image models) or as execution aids (e.g., AI-powered brush tools in drawing applications) (Chung et al., 2021b).

Hwang et al. (Hwang, 2022) further characterize how these tools apply AI models in the creative process as falling into four general categories: editors (facilitate execution of processes), transformers (aid in changing existing content), blenders (combine 2 or more content sources), and generators (produce novel content). In this framing, the large-scale TTI models that this work is focused on fall into the generators category.

Additionally, research in the related field of Human-AI Co-Creation (a sub-field of Mixed-Initiative Co-Creativity (Yannakakis et al., 2014)) is highly relevant. In the Library of Mixed-Initiative Creative Interfaces (Spoto and Oleynik, [n.d.]), Spoto et al. proposed a framework to understand mixed-initiative co-creation as a process involving seven potential actions: ideate, constrain, produce, suggest, select, assess, and adapt. Muller et al. (Muller et al., 2020) extended this framework to generative AI applications, while Grabe et al. further simplified this extension and characterized four primary interaction patterns concerning GAN applications: curating, exploring, evolving, and conditioning (Grabe et al., 2022). In our work, we observe similar themes, especially the notion of users of TTI models feeling like they are using the models to explore, and acting as curators of their outputs.

Work in this field has also identified core challenges in creating human-AI co-creation systems. For example, Chung et al. (Chung et al., 2021a) identified the limited ability to control the output of generative AI models and proposed using gestural input to constrain/guide the model output. Likewise, Buschek et al. (Buschek et al., 2021) identified a set of nine challenges system designers could encounter when developing human-AI co-creative systems. In the context of TTI models, they identified challenges of invisible AI boundaries (“A (generative) AI component imposes unknown restrictions on creativity and exploration”) and conflicts of territory (“AI overwrites what the user has manually created/edited”) as particularly salient. Participants in our study encountered similar challenges of identifying what the models are capable (and not capable) of, and in building upon prior inputs to the system.

Building on this prior CST research, we examine the emergent practices and goals that have evolved in concert with the latest generation of TTI models.

Given the rapid advancement of AI models for generating images with ever more creative and novel styles, art historians and technologists have been actively discussing how to conceptualize AI-assisted art in relation to other artforms. Experts in these communities have disagreed as to whether generative models should be considered artists in-and-of themselves (Mazzone and Elgammal, 2019) (as with the famous sale of Portrait of Edmond Belamy auctioned by Christie’s in 2018 (Christie’s, 2018)) or whether they should be considered as merely a tool employed by artists. Hertzmann (Hertzmann, 2018, 2020) argues that generative AI models are similar to the camera as it relates to the art of photography: a tool the enables the art. Hertzmann further theorizes that “art is an interaction between social agents,” and generative AI models are therefore best considered an agent in this interaction. Agüera y Arcas (Agüera y Arcas, 2017) provides a similar argument with an in-depth discussion of the similarity between the emerging field of AI-generated art and photography, particularly surrounding the historical reaction of painters to the introduction of the camera. Grba (Grba, 2021, 2022) provides a framework to critically evaluate art created with a generative AI model. In this framework, he echoes arguments above, and critiques what he refers to as “the ever-receding artist”: the repeated occurrence of technologists referring to models as artists, thereby minimizing the contribution of the human who employed the model to create art. Finally, Browne (Browne, 2022) explored what it means to be an “AI artist” and proposed the framing of an AI artist as a bricoleur (building upon (Grba, 2019)), saying, “Bricolage is common to generative art, where ideas are developed through playful experimentation with existing tools and techniques.”

In our paper, we present an analysis of interviews with artists who employ a large TTI model for the generation of their art, highlighting the work of three of these artists and provide context around their motivations and goals. In our results, we find thematic alignment with perspectives advanced by Hertmann, Agüera y Arcas, and Grba above: While the models can produce surprisingly high quality output, dedicated users of these models employ the models as tools to explore specific themes and concepts. Accordingly, they have intentionally developed processes (e.g., locating domain-specific terminology) to improve their ability to achieve their individual goals.

Diffusion models (e.g., Imagen (Ima, [n.d.]) or DALL-E (OpenAI, [n.d.]a)) are trained by gradually adding noise to an image, until all of the image is completely noise. The model then learns to reverse the noising process to generate the original image. In this way, a diffusion model learns to synthesize an image from noisy images, and is capable of generating images from arbitrary “noise.” For text-to-image models, the models are also trained (conditioned) on text inputs (nichol2021glide), allowing the model to produce an image from a noisy image input and text input, where the resulting image bears a resemblance to the input text.

Autoregressive models, such as Parti (Par, [n.d.]), treat text-to-image generation as a sequence-to-sequence problem, akin to machine translation or other language modeling tasks. In the case of TTI models, the “translation” is from text to image (i.e., text tokens to image tokens).

In our study, participants used both types of models.

For the survey, participants were recruited from an internal chat channel dedicated to TTI models (where the internal chat channel has thousands of members) and internal TTI model mailing lists (with hundreds of members). From the survey respondents, we identified interview candidates who had reported having created artwork over more than ten sessions and having spent more than five hours in the previous week using a TTI model. To create a pool of participants, we recruited eight of these latter respondents, and further recruited three prominent artists in the internal artist community to participate. These three artists are quite visible in the internal artist community, and have shared their unique artwork collections within that community. Participants were also actively engaged in external communities, sharing knowledge, expertise, and artwork. Participants were given a 60 USD gift card for participation.

The study consisted of four parts intended to understand participants’ practices. Each participant was first asked to create an image of their choice to allow the researcher to observe their natural practices. In the second part of the study, participants were asked to reflect on 1) an artwork they were proud of, 2) a piece they found most successful, and 3) the piece that was least successful. In the third part of the interview, participants were asked to reflect on someone else’s work by examining only the prompt, and specifically asked to either improve or change the prompt in their style. In the last part of the interview, participants were asked to discuss envisioned uses for these text-to-image models.

Interviews were conducted remotely. We recorded the shared screen and automatically generated transcripts for the interviews. We constrained our focus to interfaces that only use text prompts as input to the models. In addition, some participants voluntarily shared their collection of generated artwork after the interview.

For the qualitative analyses, the authors analyzed the video transcriptions and also noted comments on participants non-verbal interactions. The final corpus of automatically generated transcripts was 164 pages (60614 words). The first two authors each reviewed the transcripts data independently, looking for ways of explaining the artistic practices (Miles and Huberman, 1984). In this process, the authors separately analyzed each transcript to extracted salient themes, and independently generated hypotheses and points of discussions (Braun and Clarke, 2006). Using these data, all authors participated in two rounds of interpretation sessions to arrive at the primary themes reported in this paper, and resolve any discrepancies and disagreements. During the interpretation sessions, authors also analyzed the prompts and the images created by the study participants to identify unique artistic styles and practices. These sessions were inspired by existing analysis practices from qualitative media analysis (altheide2012qualitative).

Shai Noy is a software engineer with no training in the visual arts or design. However, they have produced thousands of images with TTI models. The styles developed by this artist include “super macro photography” images (i.e., extremely close-up views of objects, Figure 2(a)), and fashion (dresses, suits) made out of unusual materials, such as wood, grass, brick, or ice (Figure 2(b)).

Elements of both discovery and community were emphasized as rewarding for this artist, such as being the first to explore particular concepts and the ability to share discoveries: “Everything is more fun when you can share it” and, “Art doesn’t live in a vacuum, nobody starts from scratch, everything is based on something else. I am proud of being able to recognize the potential” (A1).

Irina Blok is a designer who has done visual work for many years using stock images and applications like Photoshop. This artist’s styles include origami dancers (Figure 3(a)) and reality “mash-ups,” such as sliced produce that contains different textures internally (e.g., a sliced head of cabbage that reveals a cross-section of an orange, Figure 3(b)).

Driving these styles is a passion for developing “impossible objects, something we haven’t seen before, mashups” (A2). They also seek to create images that differ significantly from the images it was trained on: “The further away the generated images are from what it was trained on, the more the satisfaction” and, “they’re [the resulting images] also defying the rules of its training set, like defying […] gravity. [..] A creative prompt breaks that” (A2).

Importantly, this artist’s output also includes prompt templates that describe a particular, parameterized image or visual concept, such as this prompt for generating a house: Audacious and whimsical fantasy house shaped like ¡object¿ with windows and doors, ¡location¿. These prompt templates are carefully crafted to reliably produce pleasing results for others. We describe this concept more fully in a later section.

Notably, when working with the text-to-image models, A2 considers the model to be akin to an artist itself, with themselves the art director. In this relationship, A2 acknowledges a certain lack of control: “[I] don’t have full control, and there’s beauty in this” (A2). At the same time, they also consider the model to be a tool: “[The model] is a brush […] you just learn to speak its language” (A2). With these dual views (model as an artist, model as a tool), they note that “the hardest part is [the] conceptual aspect, being skillful with [the] prompt” and that “it’s a thought exercise, it’s not a visual exercise” (A2). It’s really about how to make people think like an artist” (A2). In this latter sentiment, this participant was specifically speaking to the need to think like an artist in formulating a prompt, as opposed to formulating a prompt as if one was talking to a machine.

Dan Smith comes from a background in visual media, and is partially driven by the desire to deliver a message about the climate crisis, in order to facilitate change and awareness: “[I’m] not just having fun … but [actually] making something that has some power” (A3). In working towards this goal, they want to make “something that you look at, and it makes you feel something” while also making “something that people would want to look at” (A3). A3’s image styles align with these overall goals: They have a number of pieces that put nature in “situations where you wouldn’t see it” (Figure 4(a)) and have created numerous “hybrid animals” (Figure 4(b)). This particular style—hybrid animals—is also inline with a motivation to create images that “would be hard to […] visualize or create, if you were […] a really skilled Photoshop artist” (see Figure 4).

A3 also heavily considers the quality of the image when assessing its outputs: it must have near expert-level composition, photorealism, and/or artistry. When describing the artwork they created and how they created them, they noted, “I think the ones that I pick out are […] standouts for various reasons, and just […] like what I said, […] composition photorealism, artistic quality” (A3). Consistent with their emphasis on overall image quality, they have found ways to address undesirable outputs. For example, in generating images that include animals, they found they needed to adopt a specific strategy to create aesthetically pleasing images: “I would do ‘Tall Grass’ a lot because early on I discovered that limbs and fingers and paws can get a little wonky” (A3). The “tall grass” addition was their creative strategy to hide feet or paws and suggests an understanding of model limitations, but also a sense of how to cope with these limitations.

One of the primary motivations for interacting with the models was that participants found them fun—the models’ output quality enabled people to feel creative, and they were generally interested in interacting with this new class of model. Some people also noted that the model enabled them to engage in their domain interests in a new way. For example, one participant said that the models allowed them to explore their interest in Swiss trains with the model, while another found it compelling to try to create new forms of currency (e.g., new types of coins). One participant used it to create the equivalent of clip art for presentations: “I use [TTI] model when I need some […] clipart to use in my presentation, and [TTI] model would be amazing for that…” (P2).

In comparing the artists spotlighted above to the other participants, one notable difference in practices is that the spotlighted artists placed a particular emphasis on exploring specific themes in depth (e.g., origami figurines). The other participants did not pursue concepts with the same rigor or depth.

For some participants, the prompt itself was part of the overall art, and thus worthy of attention. For these participants, it was important to “[create] aesthetically pleasing images” and “[develop] art concepts” that were inherently tied to prompts. We detail these motivations and behaviors below.

While generating aesthetically pleasing, “glitch-free” images was a common goal of the creators, other goals were also present in their practices. For some participants, the prompt itself was part of the overall art, and thus worthy of attention 1) on its own, and 2) as it relates to the image: “It’s part of the aesthetic” (P3), where the prompt is “like a title of the piece, but you don’t get to choose it independently” (P3). Hinted at in this last comment is the notion that while a prompt could also serve as a title of the art piece, there is a clear dependency on, and perfunctory role for, that prompt as well: The prompt serves as the source material for the model generating the resultant image. Given this dependency, finding a prompt that produces the desired result and that can serve as a title for the piece can be challenging, but rewarding when it happens.

This same artist (P3) further described the prompt’s role as “communicating the image, and the idea of the image, and how I got it all at the same time” (P3). This quote sheds light on how art with TTI models can, in some sense, be considered multimodal (text and image) for both the artist and viewer: The prompt and the final output combine into a single, mutually-reinforcing, art piece. Seen in this light, one can consider the prompt as art itself—a well-crafted prompt that creates a compelling image but also accompanies that image, saying something about the image. See Figure 5a for an example of “prompt as art,” as well a prompt that wasn’t able to achieve this same level of aesthetic (Figure 5b).

Diving deeper into this theme, we observed that the artists believe the “concept” is a critical component of an artwork. In an internal blog article, A2 describes, “There’s a common misconception art is largely about drawing and painting skills. Art is not only about how something looks, it’s about what it says, it tells a story, and has a concept. Art can surprise, provoke, teach, delight and inspire. Art is not just about drawing, art is a way of thinking.” A1 also suggested “the unit of shareable artwork is not necessarily a specific image but maybe it’s the whole exploration of the concept of those images.”

In pursuing “prompt as art,” we observed participants impose different constraints when creating their prompts, with some wanting it as descriptive as possible, while others attempting to make it as simple as possible. One participant even discovered delight in their accidental discovery that their “random string” produced beautiful output, and named that prompt as their proudest achievement. In this latter case, the pride comes from the joint pairing of the random string and the beautiful output—without the context of the random string, the image has less value, as the random string reveals an unexpected feature of the model.

In a spirit similar to “prompt as art,” five artists sought to produce prompt templates. A prompt template is an image description with “slots” for someone else to fill in. For example, we previously noted this prompt template created by A2: Audacious and whimsical fantasy house shaped like ¡object¿ with windows and doors, ¡location¿ (see Figure 6 for example outputs of this prompt template).

These prompt templates leverage the capabilities of the models as well as the lightweight, accessible features of prompts. More specifically, when an artist identifies a compelling composition, they can create a text prompt that allows others to create a similar composition, but with their own unique customization to it. The ease with which the templates can be shared also introduces social motivations for producing and distributing templates (e.g., to participate and contribute to a larger community of practice). We elaborate on these points below.

Prompt templates have a number of key features:

• •

  They are tightly coupled to and represent a particular artistic concept, vision, or composition (such as a “whimsical fantasy house”). These features make prompt templates conceptually richer than words or phrases used to specify stylistic characteristics of the image (e.g., “35mm” or “watercolors”, which may be used to produce a particular effect, but don’t specify a larger composition).

• •

  Others can make use of prompt templates by filling in the blanks. The richness of the models means that the templates guide the overall generated image, but users’ unique input can yield a diverse variety of outputs.

• •

  There is the intent for the prompt template to provide consistently high quality and delightful output when used by others.

Unpacking these concepts in the context of the example prompt above, the skeleton of this prompt embodies a particular (visual) concept: A fantasy home in a given location. Someone making use of this prompt can customize it through two key variables: A shape for the house and a location for the house. While these are seemingly simple variables to customize, the template gives the user great flexibility in terms of the final outputs produced by the model: Any number of shapes can be provided in any number of locations (in fact, the user is free to substitute any text in the slots they wish).

Simultaneously, the prompt cues the model to the types of output to produce, as well as details to guide the generation. The phrase “audacious and whimsical fantasy house” defines desired attributes of the house, while the specification of “with windows and doors” provides additional details that should be included in the generated house. These details in the prompt help increase the reliability of the model output and reduce the likelihood that the downstream user of the prompt template needs to experiment further with the overall prompt structure and content to obtain a good output.

A noteworthy characteristic of these templates and the text-to-image models is the rich interplay that results between the original template and the phrases the user substitutes in the template. For example, a particular choice of location can profoundly influence the resulting house generated by the model—the model does not simply generate the same type of house and situate it in a different location. Instead, the choice of location can directly interact with the other parts of the prompt. For example, creating a strawberry house in the “countryside”, “Paris”, or “Tokyo” yields qualitatively different outputs for the house, with the house style meshing more naturally in the chosen location (e.g., having styles of windows more typical for a European building for the house in Paris, and doors more typical of Japan for Tokyo—see Figure 6). These interactions between the template and the users’ choices enable a diverse variety of outputs that allow a user to explore a wide range of ideas.

To produce prompt templates, one participant described a process whereby they would 1) input a prompt, 2) identify high-quality outcomes, and 3) rewrite the prompt to try to produce that same outcome again. This process could involve several iterations until they get a reliable prompt. Once a prompt is working, participants would sometimes remove content to make it more succinct. For example, they would first emphasize a specific characteristic (like “high resolution”) by applying many descriptors representing a similar effect (such as “high res”, “DSLR”, “crystal clear”, “photo realistic”), then start removing the repeated qualifiers in the prompt until the prompt could reliably generate similar outputs to the original, verbose version.

As with the notion of a “prompt as art,” the creation of these prompt templates can also be considered an artistic outcome in and of itself: The prompt author must first develop a compelling concept, then ensure it has enough capacity to enable people to produce their own unique creations within the frame of the concept. When done well, a prompt template has the quality of attracting users (because of the compelling output produced by the prompt) and continually delighting users with the outputs produced using their unique input.

Elaborating further on this notion of “prompt template as art,” a prompt template represents a particular artistic vision, with the prompt text capturing the overall design, composition, and aesthetic intentions of the author. Notably, because the artistic vision is captured in natural language text, a prompt template can be used across TTI models: Users can customize the template as desired, then tweak the prompt to produce the desired output using whatever model they have at their disposal.

A common goal of many artists was to discover new capabilities of the model that others had not yet found. As one participant put it, they tried to “break” the model, pushing it to its limits. As with prompt templates, these newly discovered capabilities are often shared so others can apply the concept in their own images. For example, an artist may identify the ability of the model to produce physical sculptures out of unusual materials or to create origami-like figurines (as A2 did). Once a concept and ability have been identified, others can build on the core idea and create their own images or permutations. In particular, participants felt proud when they “made the model do X (new concept)”, especially when they discovered a model capability for the first time in the community in which they heavily engage, as the discovery could influence the discourse in the community. We expand on these points below.

new words to steer the model in new directions. For example, one artist mentioned referencing architectural blogs to learn that domain’s vocabulary so it could be applied to their prompts (A2). Another artist mentioned turning to a thesaurus to enrich their vocabulary for the prompt. As one concrete example, A1 (the “Explorer”) likes to employ “unusual words” to steer the model, such as “intricate” instead of “detailed”, explaining: “If you choose common words, then you get a bit of an uninspiring result quite often. But if you use something a bit more unusual, then you really narrow down […] the set [of images], and you’re going to get into things that use this less common word” (A1).

What’s noteworthy here is that in the pursuit of novel imagery, creators would carefully research and choose words to produce specific visual outcomes: What you say and how you say it is critical to producing high-quality imagery with text-to-image models, requiring TTI users to enrich their natural language vocabulary in order to develop and skillfully execute a unique visual language.

Driving these practices of discovering new capabilities was a clear desire to push the model away from “average” outputs and get it into more unique spaces. In this sense, the artists are navigating and charting the vast latent space of the model and sharing back the most interesting places discovered. When a new, unique output space was discovered, artists expressed a certain satisfaction in having discovered that space: “ A good rule of thumb is to be more descriptive than not […] if you see that something is lacking, then you try to add more descriptors that will encourage the model, in this direction […] sometimes you just have to reword a sentence or move something from one sentence to another […] it’s mostly like binary search” (A1); “ I think your choice of vocabulary is very interesting because you want a large tree. But then, instead of saying a large tree you say mature tree […] Where do these […] choices come from? They just come from trial and error” (A2).

In seeking unique spaces, one participant (P8) expressly hunted for interesting imperfections; instead of avoiding glitches, they sought glitches. For example, this participant found one of their prompts created imperfections in its output for “a hybrid of a clock and a snail on an infinite mirror. Steampunk. DSLR photo. astrophotography” (Figure 7). Here, A4 explores the imperfections of the infinite mirror: it’s “technically a failure but still amazing” (P8). P8 also discovered that the model sometimes has trouble generating the backs of things (like cats) and developed that behavior into its own art style (Figure 7).

One question that often arises on the topic of human-in-the-loop, AI-generated art is one of creativity and originality—how much can be attributed to the AI versus the person (Hertzmann, 2020). Our participants also struggled with this question, with some seeking to ensure their outputs were novel. Participants wanted to validate the originality of the artistic concept, but could only check the originality of the artifact. For example, one participant described a practice of using Google’s image search on compelling outputs to ensure originality: After producing a creative output, they use image search to search for that image (or similar images) to ensure it is, in fact, original. This participant mentioned they do this in part because they are sensitized to the fact that machine learning models can sometimes memorize portions of the training data. Given this, they want to ensure their output is not in the training data.

In our interviews, the use of thesauri and domain-specific blogs (e.g., architectural blogs) illustrate the desire of artists to identify unique terms that help them produce results that rise above the average. Embedding these search capabilities directly into the tooling could be useful (e.g., quick access to a thesaurus or an embedded search engine). Pushing this idea further, there may also be opportunities to make use of large language models (LLMs) and/or the TTI model’s training data to help surface salient terms for a given topic. For example, an LLM may be able to generate terms-of-art for architecture using a prompt like this: Here are some terms specific to describing the architectural design of a house: 1). In our own test with an LLM ⁵⁵5Specific model anonymized for review, this prompt yielded terms like entryway, foyer, entry hall, and elevation. To use the training data itself to identify new spaces, it may be possible to collect terms associated with a particular topic, such as “house,” then identify other terms that are more frequently found near that topic in the training data compared to the rest of the training data (e.g., using a method such as term frequency-inverse document frequency (TF-IDF)).

To help users understand how unique their word choices are, one could also visualize the input prompt with respect to how common each individual input token is. For templates, one could also show what the common terms would be for filling in the blanks to help people move beyond those common terms to find more distinct, unique inputs. One possible outcome in helping people find more novel inputs is that their inputs may lie outside of the training distribution, leading to unpredictable results. Providing feedback through mechanisms like visualizations (e.g., showing frequency in the training data) could help users better understand these types of issues should they arise.

As mentioned, there is a desire to validate the originality of the image produced by the model. Streamlining the process of doing an image search with an online search engine is one obvious way to address this issue. However, as is the case in aiding novelty (described above), there may be opportunities to take advantage of the training data itself. More specifically, in addition to external search, one could also search for the closest images in the training data to the image produced.

While the above mechanisms would help validate the outputs generated, there may also be opportunities to help validate the originality of the input. For example, one might be able to search the training data to identify which parts of a prompt exist within the training data.

Prompt templates provide a way for an artist to derive a new style, then share it with others so they can produce their own unique images. One could imagine embracing this practice and transforming a prompt template into its own first-class interface.

For example, one could imagine allowing users to provide multiple inputs for each slot of a prompt template, then generating the cross-product of all the inputs. One could also make use of the fact that the prompt text for the templates is tokenized into vectors. Specifically, by supplying two different inputs for the same slot, embedding vectors for each input could be obtained and then automatically interpolated between the embeddings to produce a spectrum of outputs. For example, for the shape of a house (in the previous house prompt template), the user could provide two inputs: strawberry and apple. The system could then produce the embedding vectors for those inputs, then interpolate between those embeddings to create a series of images that morph from a strawberry to an apple. However, one thing to keep in mind with these interpolations is they are occurring in the text space rather than image space—the model will be interpolating not between shapes (per se) but between the linguistic concepts of strawberries and apples (which may still produce an interesting morphing between these conceptual entities).

One the primary ways the three spotlighted artists (A1, A2, and A3) distinguished themselves from others was their perception of TTI models as an art medium, with a clear focus on exploring the capabilities and limits of medium itself, rather than only on individual outcomes. In this spirit, they embraced the limitation of not being able to edit the images directly, accepting the text-only input as a defining feature and characteristic of the medium. For example, A2 expressed “there’s beauty in it” when describing the prompting interaction with the models. In contrast, other participants focused more on the outcome, and expressed desire for additional features, such as direct editing of the generated images. Embracing the TTI models as-is further reinforces the idea that text prompts are part of the artwork, rather than simply a means to an end.

This observation suggests that the design implications for TTI models can be considered from multiple perspectives: prompt-only artists, and creative professionals using the models to achieve specific goals. Creative professionals with a specific design goal may require and request specific features that offer more fine-grained control, perhaps with tools to help understand model behavior. For example, features related to directly editing the generated images were among the most frequently requested features in the survey. Given that non-design experts can also learn to use TTI models to quickly demonstrate and visualize artifacts, we hypothesize that this may facilitate more active and iterative communication between design professionals and their clients, with clients more directly engaged in the creative process (as opposed to the more traditional pipeline-like design workflow). If this proves to be true, feedback and collaboration features could become more important.

Our participant pool was drawn among employees of one large US-based corporation, and does not cover other possible ways that culture, community, and collaboration might shape use of TTI models (e.g. on social media). Also, since our analysis was episodic rather than longitudinal, we do not document how artistic prompting strategies may evolve within individuals. For the interactions we could observe, observing participants’ interactions with the TTI model does not definitively indicate their conceptions of how the model works or how best to prompt it. We also acknowledge that different models behave in different ways due to their structure, training data, and the design of the interface. In that regard, we also acknowledge some of the findings might be model dependent, and are specific to the models used in the study. Similarities and differences in art forms and art practices might be observable with different models and communities. Another limitation is that while our participants are actively involved in multiple communities, we did not ask participants about their experiences with other models or other communities in depth.