Training Mastery
Custom LoRAs.
Inject specific characters, objects, and art styles into massive AI models without retraining from scratch. Here is everything you need to know.
What Do LoRAs Do?
LoRA stands for Low-Rank Adaptation.
Base models like FLUX.1 or SDXL are massive (often 6GB to 23GB in size) and take thousands of GPUs to train on the internet's data. You cannot easily teach them a new concept (like your face, or a specific product) by just prompting.
A LoRA is a tiny, lightweight mathematical patch (usually 50MB to 500MB) that slides into the base model during generation. It forcefully steers the model's existing knowledge to generate exactly what you trained it on.
Character LoRAs
Train on 15-20 photos of a specific person's face to generate them perfectly in any scenario.
Style LoRAs
Upload 30+ images of a specific artistic style (e.g. 1990s Anime, Tarot Cards) to force the model to render in that exact aesthetic.
Product LoRAs
Train on a specific sneaker, can, or packaging to generate infinite photorealistic lifestyle shots of that product.
Hardware Requirements
Training a LoRA requires significant GPU VRAM. You can train locally on an Nvidia GPU or rent a cloud instance (like RunPod).
Time Estimates
Training time scales exponentially based on your dataset size, the base model, and your GPU speed (e.g., RTX 3090 vs RTX 4090).
15 images, 1500 steps on an RTX 4090.
40 images, fp8 precision, 3000 steps on a 24GB GPU.
100+ images, heavy regularization, intense learning rate.
Using LoRAs in ComfyUI
Place the File
Move your newly trained `.safetensors` file into the ComfyUI/models/loras folder.
Load the Node
Right click the canvas: Add Node > loaders > Load LoRA. Connect it between your Checkpoint Loader and KSampler.
Trigger it
Select your file in the node. Set the strength between 0.6 and 1.0. Include your trigger word (e.g. "ohwx man") in your positive prompt.
Pro Tip: Multiple LoRAs can be stacked by daisy-chaining "Load LoRA" nodes together. However, mixing too many styles and characters simultaneously will deep-fry the image output. Keep it under 3.