This is a WF for a Flux and SDXL Lora trainer which uses minimal VRAM and puts out excellent results. It's based on:

This guy's work. I've refined it crushed it down to what's needed and made it much tidier. Added a universal Renaming/Keyword node and an INT node to govern the steps taken per iteration to make life easier and added 40 potential dataset nodes should you need em.

Resize large images to 1024 as their largest dimension.

0.0003 learning rate at 100 steps as a 'Tester' (don't be surprised if you get a black and white preview shot first time) then 450 in a 'Big Step' to get over the hump of training with 75 steps a piece after that so you can test them out.

5 Nodes are specifically able to drop their min bucket res so you can refer to folders with smaller resolution pics in those nodes.

No captioning necessary works very well without them.

I advise restarting after each training session to ensure whatever cache is clear of old info.