SDXL LoRA Trainer

Train your own SDXL LoRA on a set of your images, then drop it into any SDXL workflow.

Point it at a folder of captioned images, give your subject a name, pick how strong and how long to train, and run it. The workflow trains in stages and renders a few preview images along the way, so you can watch your subject take shape and stop at the version that looks right instead of finding out at the end.

It builds on JuggernautXL by default and saves the finished LoRA straight to your models folder.

How do you train an SDXL LoRA in ComfyUI?

Put your images in a folder with caption files, give your subject a concept name, then set the rank, total steps, and dataset resolution. The workflow runs training in four stages and saves a checkpoint plus preview images after each one, so you can compare epochs and keep the best.

Dataset folder and concept name This is the one thing you have to set. Point the dataset at your captioned image folder and pick a short name for what you are teaching the model. The captions describe each image so the model learns the right thing.

Network dim and alpha (rank) Controls how much the LoRA can learn. Default is 16/16, a good middle ground for a single character or object. Want to capture a detailed style or lots of variation? Go up to 32. Training something simple and want a smaller file? Drop to 8.

Total steps and steps per stage Set to 3000 steps, split into four 750-step stages. More steps means a stronger fit but a higher risk of overcooking. Fewer subjects and clean data train faster. The catch: too many steps and your LoRA stops being flexible and starts copying your training images.

Optimizer Prodigy Plus Schedule-Free is wired in by default and tunes its own learning rate, so you can leave it alone. Adafactor is included as an alternate if you want manual control over learning rate and warmup.

Dataset resolution Set to 1024x1024 to match SDXL. Want the LoRA to hold up across different framings? Add a second dataset entry at a different resolution to train on multiple sizes.

Validation settings The preview images use euler, 20 steps, CFG 7.5, and a fixed seed so every stage is compared on the same footing. Change the sample prompts to your own subject so the previews show what you care about. Play with these once you see how your first run looks.