EyeBaker : The Design Prompt Database for MidJourney

We at StreamZero are happy to announce the launch of EyeBaker. EyeBaker is a prompt reference database for Architects, Interior Designers and Design aficionados, particularly those who navigate the creative landscapes of MidJourney. EyeBaker is part of a research project into AI automation by StreamZero.

The EyeBaker website offers a curated database of prompts, specifically tailored to inspire and facilitate architectural and interior design projects. What sets EyeBaker apart is not just the variety of prompts it provides, but also the accompanying images that showcase the potential outcomes of these prompts.

Each prompt is tagged, allowing users to effortlessly search and find related prompts that align with their specific needs or interests. Additionally, EyeBaker enhances user experience by including links to the job IDs of sample images, enabling users to retrieve and reference the original MidJourney jobs with ease. This feature is particularly beneficial for users seeking to understand the nuances of how different prompts influence the final design.

The primary goal of EyeBaker is to streamline the creative process for its users, offering a visual guide that helps in quickly identifying the most effective prompts, thereby saving time and enhancing the quality of design exploration in the fields of architecture and interior design.

The prompts are manually sourced and are being used as input for a research project by StreamZero into the realm of ‘topic specific prompt spaces’ with a focus on architecture and interior design. The research output is being used to build better tools for prompt engineering and understanding of blackbox AI models.

What are ‘topic specific prompt spaces’ ?

EyeBaker, was initiated by StreamZero as a demonstration project that delves into the realm of ‘topic specific prompt spaces’ with a focus on architecture and interior design.

This exploration is grounded in the hypothesis that the precision of prompts in directing image generation models can be significantly enhanced through a systematic taxonomy based approach. By analysing user-generated prompts and integrating established design taxonomies, EyeBaker aims to map out the ‘language’ understood by these models.

This understanding is pivotal in developing more intuitive tools for prompt generation, such as conversational prompt creation assistants, thereby streamlining the creative process.

A key aspect of EyeBaker’s technical exploration is the automation of prompt tagging using AI tools. This involves not only assigning relevant tags but also evaluating the efficacy of these tags as determined by AI engines. The project also delves into the normalisation of tags, moving beyond literal tagging to more abstract or conceptual tags, capturing the subtle thematic elements of prompts.

Furthermore, EyeBaker explores inferred taxonomy extensions, creating new categories and tags that are conceptually related to original tags but extend beyond them. This approach requires a sophisticated AI model capable of understanding and predicting these relationships, enriching the prompt database and offering users a more nuanced set of search parameters.

The inferred taxonomy approach also leads to the development of a ‘language model’ specific to an image generation model whose training taxonomy is not transparent. This model helps in identifying gaps in the AI’s understanding of the visual space, particularly in representing specific styles. It is instrumental in facilitating the representation of underrepresented architectural styles and inspiring the creation of non-obvious or unconventional images.

Inferred Taxonomies and Their Application

The process of taxonomy augmentation involves two primary strategies: integrating existing off-the-shelf taxonomies and developing inferred extensions of the taxonomy that maintain a certain degree of conceptual distance from the source tag. Both approaches aim to enrich the prompt database and improve the relevance and precision of search and discovery processes for users.

1. Integration of Existing Taxonomies: Utilising established taxonomies in architecture and interior design is a logical starting point. These taxonomies provide a structured, well-defined framework that can be directly applied or adapted to categorise prompts. The challenge here lies in effectively merging these pre-existing structures with the unique dataset of prompts in EyeBaker. This integration must be done in a way that respects the original intent and detail of these taxonomies while also making them applicable to the AI-generated prompts and images. This process not only aids in standardising tags but also ensures that the prompts are in line with professional and academic standards in design.
2. Inferred Taxonomy Extensions: Perhaps more intriguing is the exploration of inferred taxonomy extensions. This involves creating new categories and tags that are conceptually related to the original tags but extend beyond them, capturing nuances and subtleties that might not be immediately apparent. For instance, a prompt tagged with a specific architectural style might lead to inferred tags related to the materials, cultural context, or historical period associated with that style. This approach requires a sophisticated AI model capable of understanding and predicting these relationships. The ‘degree of distance’ from the source tag is a crucial factor here — it must be enough to provide new insights and connections, but not so far as to become irrelevant or misleading.

Both these approaches to taxonomy augmentation require a delicate balance between leveraging existing knowledge structures and innovating new ones. They necessitate a deep understanding of the subject matter, as well as advanced capabilities in AI and machine learning, particularly in natural language processing and semantic analysis. The potential outcome of this exploration is a more dynamic, intuitive, and richly layered prompt database that can cater to a wide range of design needs and preferences, ultimately fostering a more creative and efficient design process.

The inferred taxonomy approach offers a fascinating avenue for developing a ‘language model’ tailored to an image generation model, especially when the training taxonomy of the latter is a black box. This approach not only aids in deciphering the model’s understanding of visual concepts but also reveals critical insights into its limitations and biases. Let’s delve into the implications and benefits of this approach:

1. Developing a Model-Specific Language Model: By analysing and extending the taxonomy used in prompts, EyeBaker essentially creates a language model that is uniquely attuned to the specific image generation model it interacts with. This model-specific language framework helps in translating user prompts into terms that the AI is more likely to understand and respond to effectively, based on its training and inherent biases. This is particularly valuable when the underlying training data of the AI model is not transparent, as it provides a method to reverse-engineer or at least approximate the model’s ‘thought process’.
2. Identifying Gaps in Model Training: One of the most significant outcomes of this approach is the identification of gaps in the AI model’s understanding of the visual space. By mapping how the model responds to various prompts, especially those that extend the existing taxonomy, EyeBaker can pinpoint areas where the model’s training is lacking. This could be in terms of certain architectural styles, design elements, or cultural nuances that the model does not represent accurately or comprehensively. Understanding these gaps is crucial for both improving the model and for guiding users on how to work around these limitations.
3. Facilitating Representation of Underrepresented Styles: The inferred taxonomy approach is particularly beneficial in addressing the issue of underrepresentation in AI-generated images. By using a building block approach, where different elements and styles are combined in novel ways, EyeBaker can encourage the generation of images that represent less common or underrepresented architectural styles. This is especially important in a field like architecture, which is richly diverse and constantly evolving. The ability to generate images that go beyond the most common or popular styles can be a powerful tool for designers and architects seeking inspiration or looking to break new ground.
4. Inspiring Non-Obvious Image Creation: Lastly, this approach fosters creativity and innovation by enabling the construction of non-obvious or unconventional images. By understanding and manipulating the taxonomy, users can craft prompts that lead the AI to combine elements in unexpected ways, potentially leading to novel and inspiring designs. This capability is invaluable in a creative field, where the ability to think outside the box and challenge conventional norms is often the key to breakthroughs and advancements.

Conclusion

In summary, EyeBaker aims to develop a nuanced understanding of AI-generated prompts and their effective categorization. EyeBaker hope to not only enhances the functionality of design tools but also contributes to a deeper understanding of AI models in the visual domain, opening up new possibilities for innovation and creativity. For us at StreamZero the project allows us to showcase the power of StreamZero’s EventDriven Automation tools both in the space of AI as well as content generation and management.