#1 – have a quality image selection
A picky selection of images and corresponding captions is the key to create a useful LoRa network. Having a good selection of hires images is preferable to having more, but less quality images. Quality over quantity is key here.
#2 – rule of thumb for captioning for style LoRa
When automatically captioning the images, ( WD14 for example ) is becomes handy to use an editor tool, to manage all captions consistently. As a rule of thumb: If you want to train a specific style, like b/w photography with a specific analog film grain, you must delete all captions, that describe the style itself. This might be in this example: monochrome, black and white, grain, film…
#3 – LoRa versions and interations
A custom LoRa does not come right out of the box pretty. You can plan several variations in setup and training data to target the best results.
lomotrx.safetensors
A pretty practical and simple entry to the field is the tutorial video by https://www.youtube.com/@AI-HowTo. Send props!