I'll look into how I can make one soon I think, I'll likely spend a bit of time next week experimenting again since I tend to do that once a month. It'll just be kind of goofy when a summary of how I figure out what works it to just follow your heart. Random numbers in things 'just works' when it feels right.

@sekoseko I use cosine with 3 restarts, lion optimizer, 10 epochs, 5 repeats, and a whole bunch of random stuff everywhere else. I've uploaded my current presets for pony and illustrious to https://75.gigafile.nu/0625-b253084309ca4763b431bde8d24af4da3 as of 16/mar/2025, it'll likely evolve from here, but this is what I currently use with Kohya_ss, the only other information you'd really need outside of that config file for what I do is to set up your folder to do 5 repeats.

For tagging I use https://github.com/Particle1904/DatasetHelpers tagging option with WDv3Large at a threshold of .2, and then using the process tags to add character or style tags that I feel I want or need.

That's about it really, I sort of change it up every month or so when I decide to get experimental, but that's where I'm currently at.

@sekoseko Cosine with restarts seems reliable with how I have it, but it's not perfect, I'd say 1/20 models it completely messes up because for some reason it's 'loss' just skyrockets rarely on a model and just wipes the lora data out on the last restart. Likely due to my learning rate being too high, so that's a starting direction to look for, for refinement, If I find a better preset I'll try to return to this thread when or if I do.

@sekoseko I tend to stick to native resolutions of images as I find them, you can use the dataset helper I linked earlier to resize them, or 'fix them' if they're gifs or strangely corrupted where they break the baking process at the start. I don't think you can normally train with gifs, but if you're lucky enough that the first frames of the gif animations are a useful image, you can convert the gif and roll with it using the resize function. I've had success as low as I'd say 6 images, but you have to be sort of clever with how datasets that low of images are, usually you can double them using Irfanview which is a free program to flip the images horizontally and effectively have double the viable dataset, not entirely sure why that trick works, but it helps. There's options built into the baking software of 'horizontal flip augmentation' but I find that having flat out a doubled small dataset with manually flipped works better than bothering with that mystery option.

I tend to use comics and doujins at their full dataset size assuming they're under 200 images, with unnecessary images pruned out such as advertisements, credits pages, and pages featuring no or little relevant art. If a dataset I've picked up has over 200 images I'll tend to trim it down to a maximum of 120-240 images, kind of depending on how I feel at the time and if I feel like if it's worth the extra baking time for a larger dataset.