The Gap Between Imagination and Visualization
In the creative process of architecture, two fundamental moments coexist: the imagination of spatial attributes and their representation. The creative translation of this imaginary involves media, languages and technologies that transform mental images and concepts into tangible ones. With the incorporation of digital technologies such as 3D modeling and the simulation of light and materials, the author's range of choices expands exponentially, allowing for an almost infinite number of visual combinations (what Pierre Lévy (1996) refers to as "potentialization"). Computational processes thus increase the repertoire available for user selection (Austin, 2015).
However, the complexity and constant evolution of 3D modeling software for realistic visualization make their smooth integration into creative processes difficult (Popova et al., 2021), limiting their application to isolated moments, generally in the presentation of consolidated projects.
Generative AI as a Bridge
Generative AI can transform this predominantly linear flow into a continuous cycle between imagining and visualizing. Studies indicate that AI tools such as Midjourney, Stable Diffusion, and Dall-E are already being used in architectural design (Luhrs, 2024). However, incorporating the imagery produced by AI into the design process still depends on image generation time, which can take tens of seconds or more. This interruption in the creative flow hinders a more dynamic interaction in which visual results could be continuously incorporated into the redesign of the initial idea.
The Deep Mirror Proposition
Deep Mirror proposes a radical acceleration between the initial image (originating from more traditional means like hand drawing, simplified 3D modeling via SketchUp, and physical models) and the AI-generated image in the early stages of creative processes in architecture. The hypothesis is that continuous feedback, without interruption, between human imagining and machinic imagining, combined with strategies for dynamically changing the prompts and image generation parameters during interaction, can result in an output similar to a three-dimensional animation.
Automated (AI-based) retention, storage, and classification by the Deep Mirror system of the generated images (for later consultation and incorporation into new processes ) can significantly expand human imagination and thus enrich creation processes in their early stages. These images, which refer to spatial qualities (ambiences, lighting, configuration of volumes and materials), could inspire later design decisions regarding functional and constructive solutions.
LLM-Augmented Creative Process
Recently, a greater participation of LLMs was introduced in the creation process: the Deep Mirror interface continuously updates itself based on user decisions. Prompts are generated from combinations of elements that update based on the user's initial choices, creating a dynamic feedback loop between human intention and machine interpretation.
Deep Mirror: Temporal Recursivity and Operative Memory
Deep Mirror reconfigures this logic by introducing the systematic recording, storage, and reactivation of AI generative data, understood as a potentialized resource (Lévy, 2000). This shift establishes a complementary operative design mechanism in which past generative states are mobilized as active components within an ongoing real-time-based design process.
In Deep Mirror, continuous video streams, snapshots, and complete sets of AI parameters and weights are treated as material traces: technical inscriptions of intensities and probabilities (Kittler, 1999; Ernst, 2013) that can be reactivated independently of their original context of production or initial user intentions. Once reloaded, these archived traces no longer function as representations of past events but become present operational conditions within the system (Manovich, 2020).
Already a central strength of the original Deep Mirror, the generation of environments through dynamic variations in lighting, materiality, and geometry acquires new dimensions within this context of temporal recursivity. Architectural atmosphere design emerges as a feedback loop in which a structured past is continuously reprocessed, enabling iterative exploration through recursive engagement with accumulated generative states.
Because atmospheric qualities resist fixed definitions and tend to emerge through sustained exposure to variation rather than singular representational acts (Zumthor, 2006), this temporal feedback logic of past–present intertwining provides a conceptual and methodological basis for Deep Mirror's relevance to architectural spatialities and atmospheres. In this framework, the past no longer appears as historical time but as an operative temporal material. Architectural design is thus redefined as a chrono-operative practice in which spatial imagination unfolds through the technical execution of stored signals, and where human intention intersects with algorithmic probability within the time-critical operations of the machine.
