Who is this guide for?
People who may understand the very basics of making a LoRa but need tips for getting higher quality results.
What to expect?
This guide is NOT a step by step tutorial.
This guide IS a distillation of various tips I've found searching through reddit comments and civit articles that may not be readily apparent.
Some very basic info is omitted because too much info dumping in one spot makes people turn their brains off and go to sleep.
Treat this guide as an index that may have specific issues you're having and their solutions. I'll update as I find more.
Tips specifically related to parameters and settings will assume you are working with a trainer that is Kohya based (onsite training for civit, or Kohya itself)
Suggested parameters
Keep it simple:
Epochs = 10
Repeats = 30
Make use of civit's test prompts to preview the images the lora can make before downloading, sort through until you find which epoch you like best. It's likely to be around the late-middle (7-8) with larger datasets, and towards the beginning (3-5) with smaller datasets
Issues
distorted faces/missing pixels
If you're getting distorted faces when using your new lora, but the settings in your workflow generate serviceable faces without the lora, you're likely over-training or using bad resolutions
--get your training images as close to 1024x1024 as possible
--set your repeats lower, the generally advised number of steps (1500 ish) assumes you have a dataset of 20-30 images. If you have a small dataset then you want your repeats to be around 20-50
shot/angle too fixed
If the generations are coming out in a particular shot (like portrait) even when you specify other shots, your lora may be too strong/biased towards that shot. You need to retrain with more varied distances/angles, but if you want to find a quick solution to make use of the lora as is: Try one of the following workarounds:
--generate the image without the lora, and then use a facedetailer node with the lora to essentially do a faceswap
--try turning the lora strength down to 0.5 and generate in batches of 4 with emphasis on your shot/angle type (put it in parenthesis). You'll probably miss with three of your gens but at least one will probably be good.
face/person looks weird/unsettling/wrong proportions
If you're seeing things like the forehead being too large or the eyes being too far apart or similar anatomy failures, your lora is likely under-trained. Increase repeats