Super-Digital Drawing: A Traceable Human-in-the-Loop Protocol for Validating AI-Generated Architectural Images

Introduction: Latent diffusion models have expanded synthetic image production, yet current image–text alignment and compositional benchmarks do not fully capture architectural consistency, disciplinary validity, or process traceability. Human–AI interaction research emphasizes feedback, correction, and user control, while architectural scholarship frames computational systems as increasingly cognitive and design-active. This study introduces Super-Digital Drawing as a methodology for governing AI-generated architectural images through a documented generative sequence.

Methods: A pilot corpus of image-to-image architectural transformations was examined under controlled variation of textual prompts, reference images, stylistic conditioning, and iterative corrections. Each sequence was recorded as intention → input → model → output → interpretation → deviation → iteration → documentation → validation. Expert visual analysis assessed four dimensions: semantic fidelity, geometric persistence, perceptual–atmospheric transformation, and process traceability. Intermediate and rejected outputs were retained to reconstruct decision paths. 

Results: Preliminary comparison showed that photorealistic plausibility can coexist with geometric drift, altered spatial hierarchies, and atmospheric resemanticization. Three recurrent conditions were identified: preservation of local visual features without global architectural coherence; normalization of ambiguous or conflictual qualities in the source image; and loss of reproducibility when prompts, model versions, intermediate outputs, or selection criteria were omitted. Conversely, documented iterative control made deviations interpretable and expert validation explicit.

Conclusions: AI-generated architectural images should be evaluated as process-dependent imaging artifacts rather than autonomous outputs. Super-Digital Drawing reframes authorship and control as the governance of generative conditions. The proposed human-in-the-loop protocol complements automated alignment metrics and provides a transferable framework for evaluating synthetic images in other high-constraint imaging domains.