Getting the most out of your sample pictures during LoRA training.

Results from in-training LoRAs (or “epoch pics” as I’ll call them for simplicity) are pictures returned to you on each step of LoRA training and (if using CivitAI’s onsite trainer) the final set of epoch pics will be the default gallery for your model. These are generally low quality due to their limited options, and most will (rightly) recommend an epoch pic be replaced quickly. That doesn’t mean you should just neglect them though: Firstly, you are paying for your epoch pics, so you might as well make them as good as possible to function as a basic showcase, you’ll occasionally get something really fun despite the flaws, and if you keep them in the back of a gallery you’ll still sometimes get buzz for people reacting to them. Having good in-training pics also gives you a better idea of what cautions you’ll need to write in the description (e.g., a particular tag as negative) without needing to waste time/power/Buzz experimenting further. The on-site trainer technically has a function that, if you leave the field for these blank, it will generate a random prompt for you. Don’t do it: The system for creating these is completely broken and will pick from your least used tags (often missing the trigger word) and produce horrible stuff.

Firstly: Specify everything. The use of base Pony (for Pony models), DIMM as sampler and just plain old not knowing what your model will make the system favor means you’ll often step into some weird stuff if you don’t keep it on rails, and most importantly you’re locked in to a prompt for 10 gens of a prompt even if the first 3 start doing weird stuff, so don’t take chances. You’ll want the normal art style, scores (if Pony), and subject classifications (“score_8, score_8_up, score_7_up, source_anime, anime screenshot, 1girl, female, solo”) most generations want. Past that you want to specify pose (even if just “standing, arms at side”) and shot inclusion (likely “full body”) to avoid winding up with the character is doing something strange. Including face expression directions isn’t need, though something simple like “smile” or “smug” and “looking at viewer” won’t hurt. Next, and comparatively unique to these over normal gens, is to include a background. Even just “simple background, white background” is far better than no background at all, as in-progress gens have a tendency to include weird stuff (like corrupted data looking static) if just given a character with no background specified at all. More complex backgrounds are possible, and recommended if you're going to have more than ~10 epoches, as the background is a good indicator a LoRA was overbaked.

The above will give you a skeleton, applicable to most character LoRAs, of something like (obviously tagging in 1boy, male focus if doing male character)

score_8, score_8_up, score_7_up, source_anime, anime screenshot, 1girl, female focus, solo, [character details], looking at viewer, smile, full body, standing, arms at side, simple background, white background

This should get you a usable first pic. But what of the second and third? For the second, assuming there's no variants (alt costumes etc.) that also need testing, I like to test how a character generates from behind. Just take your normal prompt, remove things only visible from front (eyes, hair parts like "bangs" and pieces of clothing only present on the front) and add "from behind". It’s relatively simple, but gives a lot of information on how the LoRA is functioning (especially as, if you’re not pulling from the deepest data set, you may have to accept having few, or no, pics showing the back and some LoRAs handle that limited data set better than others).

For the third, you’ve got the major stuff out of the way. My preference here for characters is trying to put them in some kind of alternate outfit. If you’ve got some kind of secondary outfit in training data, you should run that. Otherwise put them in some generic clothes that don’t overlap in tags with their normal outfit. If you’re making a fictional adult woman, nothing at all (nothing at all, nothing at all…) or the classic string “blush, embarrassed, looking at viewer, nude, naked, outdoors, onsen, towel, naked towel, steam, bathing, nude cover, partially submerged, water, bath, steam censor, wet towel, cowboy shot, dutch angle” is also an option (just remember your gallery needs to have at least one PG rated image for it to be visible to people not logged in). Alternatively, if you've got non-anime art styles in your TD with some kind of style tagging applied (such as a large volume of "lineart, settei", "manga illustration" or even just a lot of "official art" in a very particular style) you can try generating a picture with that to see how much the style takes hold and get a different view of the character.

If you know more than me on this, post your opinions in the comments. If you’ve got some question on LoRA training, ask and I’ll answer to the best of my abilities (if its something I know enough and can say enough about, I may even write an article on it).