Custom LoRA Training for Brutalist Architecture Renders | RunPod + GPT + ComfyUI Workflow
In this video, I walk you through a complete workflow for developing a custom LoRA model for Brutalist architectural renderings — from dataset preparation to testing in ComfyUI.
You’ll learn how to:
🧱 Collect Images: Use Google Image Search to compile a curated dataset of brutalist architecture examples — concrete forms, heavy massing, shadow play, and raw textures.
📝 Generate Prompts: Use GPT to automatically create descriptive text prompts for each image, ideal for LoRA conditioning.
⚙️ Train Your LoRA: Set up and train the model using RunPod on a custom training template optimized for SDXL. I’ll show you how to adjust parameters like learning rate, batch size, and epochs for best results.
🧩 Test in ComfyUI: Import the trained LoRA into ComfyUI, combine it with SDXL, and explore the creative possibilities. I demonstrate prompt variations, CFG scales, XY plots, and other key settings to fine-tune visual output.
This workflow is perfect for anyone looking to build domain-specific LoRAs in architectural visualization, exploring new aesthetics and rendering styles directly within generative AI tools.
💡 What You’ll Need:
A RunPod account (GPU instance)
A dataset of 5–20 images
Basic knowledge of Stable Diffusion and ComfyUI
(Optional) GPT or ChatGPT for generating captions