Reference-Based Iteration

This chapter investigates reference-based conditioning techniques in ComfyUI that enable stylistic and material consistency across iterative outputs. It introduces IPAdapter as a visual conditioning module capable of transferring high-level stylistic features—such as palette, lighting, materiality, and tectonic order—into new generations. Variants such as IPAdapterXL and IPAdapterXL Plus are reviewed alongside lighter alternatives like ReferenceOnly, highlighting trade-offs between fidelity, computational demand, and flexibility. A typical workflow is outlined, combining text prompts with encoded visual embeddings to preserve architectural coherence across exterior–interior views, programmatic variations, or design phasing. Complementary strategies such as fixed seeds, latent saving, and controlled noise injection are presented as methods for stabilizing iterative series. The chapter also provides best practices for balancing CFG scales with adapter weights, filtering noisy references, and maintaining aspect ratio alignment. Limitations—including the inability of IPAdapter to enforce geometry without ControlNet—are critically assessed. By framing reference conditioning as a mechanism for narrative continuity and stylistic coherence, the chapter positions these tools as essential for architectural storytelling, portfolio development, and phased design studies in generative pipelines.