I often get questions about how i make my Lora. As a reference i'll leave here what i found to be working for me. There are better in depth guides on this site: Please have a look at them. I started out with this one.
This applies to concept and style lora with photorealistic datasets. For anime character lora look for another guide.
1. Pick a clear visual concept for your lora.
2. Get a dataset of at least 50, better 100 images. (Fewer images can work well too particularly with flux, but if you wanna have a flexible model more is better (with diminishing returns)).
2a: Make sure your dataset is consistent and actually has a clear representation of the concept in every image. Delete the rest.
2b:Crop your images to a size of at least 1024x on the longest side (if some are lower thats ok). only if the image composition allows it i try to crop to a multiple of 256 on the longer side (so 768x, 1024x, 1536x, 2048x),
3. Tag according to your base model (i use coqvlm for flux with wd added sometimes, wd for everything else) - see software list -
4. Don't worry too much about settings. the standart settings on the civitai online trainer are a good starting point for most concepts. I aim for 1500 steps for pony 2000 steps for flux most of the time and then select the best epoch. (the only other Settings i mess around with are: learning rate (sometimes lower lr + more steps seem to yield better results) + dim/alpha.
5. If the concept fails: work on making the dataset more consistent, do some handtagging (also: try to be as consistent with handtagging as possible).
Addendum:
Software Recommendations (for Windows)
image editing - Krita
relatively fast manual cropping - Imageglass
duplicate image removal - Czkawka_Gui
batch resizing and editing - XnConvert
tagging and batch editing - DatasetHelpers
autocaptioning - Taggui
further hand tagging - BooruDatasetTagManager
I'll just leave this here, maybe i find time to update this later. If you want me to expand on something: ask away in the comments.
Cheers!