First of all you need a 3 distinct groups of images Character A , Character B and character A+B. I would recommend at least 50+ of each. Second of all you will need to do extensive tag pruning as the way you are currently tagging (1girl, pink hair, etc) will confuse SD and it will not know which attributes belongs to who. So you must prune all intrinsic attributes(eyes, hair, 1girl, etc) so they are folded into the character triggers. You do this as it you were making two individual character loras. Afterwards the A+B group must be tagged with the 3 triggers, making sure to prune any tags alluding to the number of people. Basically you are stuffing two character loras and a concept lora(two people) into one single lora. So at the end just merge the datasets, add some tweaks for extra outfits(if any) and that's it.

If you have any doubts you can use my dataset as an example, I normally add it unless it has too many screencaps and i fear to be sued in the future.

@TomLucidor I'll probably add it to my guide https://civitai.com/articles/138/making-a-lora-is-like-baking-a-cake later.

For extra outfits as long as you don't over train them they should be able to changed with relative ease. If they are simple and common enough, just replacing them for a trigger(for example outfit1) and pruning the individual parts(black shirt, blue pants, red belt, etc) should work. If the outfit is more complex, after pruning, I prefer to tag a custom trigger made of the individual parts to use some of the contamination from the model to boost it(instead of outfit1 use Black_shirt_Blue_pants_Red_belt). The second technique allows to lower the amount of repeats needed lowering the risk of it burning. You can check my guide for a more in depth, but essentially it is like a venn diagram and you must tweak the dataset so SD understands the difference between each trigger.

For the dataset size, you can say it matters and not at the same time. I mean supposing all images are of the same general quality, you could do 50 images at 10 repeats or 500 images at one repeat. Both will come out ok but the second will be much more flexible, and if using adamw as an optimizer it will require a bit higher learning rate or an extra epoch. Prodigy on the other hand should be able to compensate without much issue. Do remember to keep the total steps balanced between both characters For example if you train characterA for 500 steps per epoch and outfitsA1, outfitsA2 and outfitsA3 for 200 steps per epoch. Do the same for characterB and try to add outfitsB1, outfitsB2 and outfitsB3 with the same amount of steps.