ComfyUI Tutorial for Beginners: LoRAs, Style Transfer & ControlNets (2026)

Module 3 of Comfy 101 covers style and creative control in ComfyUI. This tutorial moves beyond prompt iteration and into the tools that give you direct control over your output: LoRAs, LoRA stacking, LoRA training, style transfer, and ControlNets. What this tutorial covers: LoRAs (Low-Rank Adaptation) — how they attach to a base model to change its output without retraining, why a LoRA must match its base model family, and where to find them on Comfy Hub and Civitai. We walk through the Inflate LoRA on Qwen Image Edit, then move to Stable Diffusion to demonstrate model strength and CLIP strength, and how stacking multiple LoRAs at different weights produces a custom style. LoRA training — building a dataset, writing trigger word captions, choosing a step count, running the TrainLoRA node locally, reading the plot loss graph, and testing the result. Includes guidance on appending to an existing LoRA and on sourcing training data responsibly. Style transfer — two approaches compared: a prompt-driven img2img method where denoise strength controls how far the image moves from the original, and an IP-Adapter method that reads style directly from a reference image with no description required. ControlNets — using DWPose, Depth Anything V2, and canny edge preprocessors to set composition before the prompt is read, with a side-by-side comparison of what each map preserves. The module closes with a practice exercise and a troubleshooting section covering weak LoRA effects, overcooked stacking, mismatched ControlNet preprocessors, and noisy style transfers. Click the link to run the workflow: https://links.comfy.org/4gsBzc2