Training and Using Custom LoRA Models

This chapter introduces Low-Rank Adaptation (LoRA) as a parameter-efficient method for aligning generative models with office-specific architectural styles. Rather than retraining entire diffusion models, LoRA inserts lightweight adapters into targeted network layers, enabling fast, memory-efficient fine-tuning without overwriting base weights. For architects, LoRA offers a means of embedding portfolio identity, ensuring stylistic consistency across speculative outputs, client proposals, and academic studios. The chapter outlines dataset preparation, emphasizing curated collections of 20–200 images with coherent resolution and style. Training pipelines such as Kohya Trainer, HuggingFace Diffusers with PEFT, and ComfyUI-compatible loaders are reviewed in detail, alongside strategies for combining or blending multiple LoRA adapters during inference. Architectural applications include competition submissions, client-facing prototypes, and portfolio continuity. Risks such as overfitting, stylistic inbreeding, or unauthorized training on third-party imagery are critically assessed, with mitigation strategies grounded in dataset diversity and transparent documentation. By reframing LoRA as the architectural equivalent of custom rendering templates, the chapter positions this technique as a practical mechanism for asserting authorship, continuity, and identity in AI-augmented workflows.