​
Training (Finetuning)

To train LoRA, add some images to a folder named data. The author recommends about 7-10 images for good results.

rubbrband launch lora

# data folder should contain 5-10 images you want to finetune. 
# you should use jpg or png files. 768x768 images are recommended.

rubbrband train lora --dataset_dir="./data"

For reference, training on an RTX 3090 takes ~20 min for 1k steps.

​
Inference

After doing a training run on LoRA, run the rubbrband eval command.

rubbrband launch lora
rubbrband eval lora --input-prompt "a man on the moon in style of <s1><s2>"

​
Where are my outputs?

To view the output image on your local machine

rubbrband copy-from lora /home/engineering/samples/output.jpg ./data

To copy the checkpoint safetensor file onto your local machine, run the following:

rubbrband copy-from lora /home/engineering/output/final_lora.safetensors ./data

​
Sample Image Outputs