sometimes you gotta boost the strength on deep pentration for it to take, check out my eighth sample image
```
(deep penetration:1.3), 1girl, 1boy, pov, missionary, on back, spread legs, arms up, torso grab, thigh grab, grab, blonde hair, long hair, medium breasts, simple background, black background, completely nude, facing away, arms up, <lora:concept-tip-mid-base-vaginal-v1:0.7>

Negative prompt: (worst quality, low quality:1.4),

ENSD: 31337, Size: 512x768, Seed: 738139466, Model: AOM3, Steps: 30, Sampler: Euler a, CFG scale: 5, Clip skip: 2, Model hash: d124fc18f0
```
I think I need to retag, a problem with the data set was that basically every image is tagged with some variation of vaginal and penetration, so I think that confused it a bit, but you can get around that in generating by increasing attention to the key words, like imminent, penetration, and deep penetration, or by not using vaginal in your prompt.

@ShonanFM

I'll think about making a full guide, but to get you started.

All I really do for these is grab a whole bunch of images that fit each stage, and make sure to tag them appropriately.

What counts as appropriate is something I haven't entirely figured out yet.

But what kinda worked for this one:

Say you wanted to do a lora on smacking or slapping. You'd want images appropriately tagged for 'about to smack,' 'smacking,' and 'smacked'.

You would want to remove all smack or slap related tags, other than the ones you add, so that you don't confuse the training. I typically use the tags that are on whatever booru I grab the images from, plus I run the tagger script on them, and then, I remove any tags that don't fit the images the way I want. In the smack lora case, you'd want to remove all smack related tags, then add your stage tags... the 'about to smack' 'smacking' and 'smacked' ones.

You want each concept to be in their own folder in your img folder for training, with the count set to give them the same total reps. So say you have a hundred about to smack images, 50 smacking, and 25 smacked?
You'd have something like

10_about_to_smack concept

20_smacking concept

40_smacked concept

That would give you a thousand total reps each, and then from there I usually run it for 20 epochs, which is a lot, but kinda necessary so that you can burn it in enough that the transitions between frames are consistent.

Then, if you lucked out, and your smack lora turned out consistent, then you can do like I did for this for the animation.

You would write prompts frame by frame that transition between the concepts you trained into the lora,

something like:

prompt 1:about to smack,...

prompt 2:about to smack, smacking,...

prompt 3:smacking,...

prompt 4:smacking, smacked,...

prompt 5:smacked,...

Note: if there's an asymmetrical aspect, as there would be for slapping, then it's important NOT to use the flip training augmentation, instead, you need to manually flip all images so that their oriented in the same way, ie, all are slapping from right to left.

@ShonanFM That is something I haven't been able to do consistently. But I think the way to do it is by having more of a certain type in with that tag. ie. if you have 10 images for deep_penetration, you might include 2 images also tagged deep_penetration that are actually mid_penetration, that way increasing penetration on deep_penetration makes it more likely to give deeper images.

Or not? I haven't actually gotten that to work consistently. On this one for instance? It works on the deep_penetration tag, (deep_penetration:1.5) is deeper than (deep_penetration:1.0), sometimes, but it doesn't really work on any of the others.

I had actually tried training a version on nothing but 'deep penetration' but with the repetitions set to let deep penetration's attention alone decide depth, but that didn't work at all.

That's the main reason I transition the tags the way I do.

@ShonanFM My samples are generated with AOM3, but the training is done on NAI.

it can be, but it's little gacha, you gotta find the right seed for it.