Kyuuuuiiiinnnn... I wanted to try my hand at making one of those mass produced 30 image dataset throw away LORAs. The result is exactly what i expected. It is stiff, weird as hell, burns easily and behaves randomly without a lot of extra tagging. So well... i wanted to experiment but didn't want to ruin a pretty lady so... I sacrificed a face up Kogarashi in defense mode!
Got ok-ish results with: <lora:KamenNoMaidGuy:0.6>, KamenNoMaidGuy, solo, standing
I cleaned the dataset but seems a lot of garbage remained due to this being a low effort lora. So check your Negatives!
Took me about an hour to clean the dataset, upscale what needed it and do the tagging. Which is very little in comparison to my normal LORAs.
Anyway I guess the conclusion is clear, at least to me. Training with less than 100 images works best with little to no tag pruning as to prevent overbaking. If you prune too much you increase your chances of overbaking.
Therefore Low effort small dataset LORAs are harder to use and require more verbose prompting, because if you consolidate the tags you increase the risk of overbaking.
Surprisingly the laziest approach seem to be the actual best method for small dataset loras. Zero tag pruning and leave the verbose prompting in the hands of the user. They might actually do better without an actual trigger.
Well regardless, I will add the dataset and tag summary. As I said this was experimental. Negative results are still results!