ComfyUI tutorial: LoRAs, style transfer and ControlNets

This beginner ComfyUI lesson covers ways to control image generation beyond changing prompts and seeds. The speaker demonstrates the Inflate LoRA with Qwen Image Edit, then uses Stable Diffusion to explain model strength, CLIP strength and stacking multiple LoRAs. Each LoRA needs a compatible base model family and may require trigger words. Holding the seed fixed helps compare changes in weight.

The local AI training example uses DreamShaper XL to generate 30 glass-cat images and trains a LoRA for 3,000 steps. The speaker reports that the run took about 30 minutes on his Blackwell GPU; this is an example, not a hardware requirement or a general timing estimate. He says the demonstrated training workflow requires a local desktop instance or a separate service because Comfy Cloud lacks the training node. The lesson also covers captions, saved weights, loss graphs and continuing an existing LoRA with more data.

Two style-transfer approaches follow. Prompt-driven image-to-image generation uses denoise strength to control departure from the source. IP-Adapter takes style information from a reference image through CLIP vision embeddings. The ControlNet section compares DWPose skeletons, Depth Anything V2 maps and canny edges. In the examples, these preserve different aspects of pose, object placement and outlines. Troubleshooting focuses on weak LoRA weights, excessive stacking, mismatched preprocessors and conflicting style references.