Wait for base z image model to drop. You are not supposed to train loras on destilled models but people train lora on destilled z image anyways because they are impatient.
As for image quantity a general rule of thumb in the past is that an AI model needs to see an image / pixel pattern 100 times to recreate it.
At the 10,000 training steps , using repeats of 5 , rule of thumb says at most you can use 10,000/(5*100) = 20 images.
Repeats are like test swings at a golf course to find the best direction before swinging.
At batch size 4 , you can do 4 images at once per training step.
So 10,000 computation steps are 40,000 training steps at batch size 4.
Meaning the max count by rule if thumb at repeat 5 for 40,000 training steps is 40,000 / (5*100) = 80 images at batch size 4.
Optimal learning rate for Z image base is still mystery but would assume rank 32 alpha 16 at LR = 0.0004 (three zeros) at cosine with repeats is worth trying.
Coverphoto from the 'Monstress' comic from 2015-2017 , kinda thinking of making Chroma lora for it.
I'm not gonna read the entire volume for it tho. Instead I use deeppanel extractor https://huggingface.co/datasets/codeShare/lora-training-data/blob/main/Comic_Panel_sorter.ipynb
Github : https://github.com/pedrovgs/DeepPanel
And then sort rge extracted paneks by CLIP aesthetic (available in notebook).
So for those eager to train Z image , would suggest spending time prepping datasets / ideas for whenever z image bade is released.
/ Cheers
