Training requirements vary with your project. how big are your images, how many are there, are you using an optimiser to reduce load (there are 15 different ones), what batch size, what Learning rate, how many Epochs? all of these factors influence the answer to your question, so the answer is meaningless.

"This piece of string here is 9 inches long."

I want to help you with a useful answer, however the answer to your questions will not help you.

Instead i'll give you some basic information on this evaluation dataset, that was used to train this Lora.

There were 11 images of 1024x1024, each image features the same character in a different pose, drawn in the same style with a similar aesthetic, however some inconsistencies in the character art do exist, so providing room for stylistic approximation.

Each image was paired with a matching filename text file, containing the tag for training:
"AssassinKahb style a demonic looking skeleton woman holding a sword with red hair"

This caption is historic, was used to train baseline checkpoints and lora's in SD1.5, SD2.1 and SDXL
I used an RTX 4090 24gb GPU, at batch 1. It took around 40 minutes to train, used up at least 12GB VRAM

I used ADAFACTOR with default settings to train.

that's the best i can do, but it only gives you hints, because your mileage may vary with your own project ;)

@tscott65 i spent more time training SD2.1 and SDXL, so i rarely worked with 1.5. It can depends on a lot of factors, however, considering the fact you get 1024x with SDXL, i have to say SDXL takes less time. it's hard to fairly compare them owing to SD1.5 smaller base dimensions. As an example, i used to train 1.5 at 768x768 and 2.1 at 1024x1024. In SDXL you can train 1280x images, although aspect bucketing means they are likely resized automatically. it's really tough to call a fair comparison.

this particular study compared SDXL and Stable Cascade, using the same dataset and settings.