Train a LoRA in ComfyUI

Channel:

SBCODE

Subscribers:

26,100

Published on August 10, 2025 1:45:30 PM ● Video Link: https://www.youtube.com/watch?v=DW9pyLH1u-8

Duration: 0:00

16,080 views

496

Channel Membership :

Documentaion : https://sbcode.net/genai/train-lora-node/

This video shows some experiments of using the "Train LoRA" node in ComfyUI.

At the time of creating this video, there was limited public knowledge that I could find. So I created a video about my experments.

I generated several combinations of LoRAs using Rank 8 or 16, and also using 1 step per training image, 32 steps per training image and 64 steps per training image. I had 10 training images at 1024x1024 pixels.

The Rank 16, with 640 steps was the best LoRA I generated.

I imagine that I could pruduce an even better LoRA using Rank 32 and 128 steps per training image.

To generate the training images, I used Flux Schnell, IP-Adapter, a detailed prompt, and some varying, or none openpose images.

My FLUX IP-Adapter workflow produces images with much better style consistency than any of my LoRA tests did. So I don't expect to use LoRA in the future.

See more info about my FLUX IP-Adapter workflow at https://sbcode.net/genai/flux-ipadapter/