Before During After
Usage
Keyword is sequential, but you don't need it.
What you do need to do is to include tags for all three frames.
Generate at 1280 x 512 for best results. Can do others? 1280 x 1024 will get you some massively long dongs, and some truly painful looking deepthroats, so maybe that's what you want?
The idea here is that you get three frames, one before, one during, and one after.
Typically before means either completely before, ie, you can see the whole hog, or it means that the lady has just started, during, is during deepthroat, usually ejaculation, and after, is the nice satisfied look of a lady (or man) that knows that she (or he) has done a good job.
Check out the samples for examples on how to use; works with pov, and from side.
Training data!
Heavily on mizumizuni, with sets included from dagashi and kashu_(hizake)
This can do scenes other than deepthroat, but it's pretty gacha.
Examples:
CivitAI keeps borking things, so hoping that this description doesn't get borked?
Here's a couple example prompts.
sequential, 1girl, 1boy, pov, deepthroat, black hair, long hair, cum in throat, before fellatio, after fellatio, large breasts, nipples, cum in nose, large penis, tears, mature female, streaming tears, mizumizuni, freckles, <lora:concept-before-during-after-v1:0.7> Negative prompt: (worst quality, low quality:1.4), bar censor, bad hands, multiple penises, Steps: 12, Sampler: Euler a, CFG scale: 5, Seed: 2527336060, Size: 1280x640, Model hash: d124fc18f0, Model: AOM3, Clip skip: 2, ENSD: 31337
sequential, 1girl, 1boy, from side, deepthroat, black hair, long hair, cum in throat, before fellatio, after fellatio, flat chest, cum in nose, large penis, tears, mature female, streaming tears, dagashi, freckles, school uniform, skirt, kneeling, penis on face, imminent fellatio, <lora:deepthroat-sequential-v1:0.7> Negative prompt: (worst quality, low quality:1.4), bar censor, bad hands, multiple penises, Steps: 12, Sampler: Euler a, CFG scale: 5, Seed: 1801869598, Size: 1280x640, Model hash: d124fc18f0, Model: AOM3, Clip skip: 2, ENSD: 31337
Description
First version, I might retrain to get the view size more consistent.
FAQ
Comments (11)
It's an anime lady, so it triggered moderator review, so I think that you're unable to see the actual images this generates.
This generates three images, side by side, and then you can cut them manually, or by script, to make gifs, like the ones I posted.
I just uploaded a script to my 'animation helper scripts' page, to help turn these into gifs.
Warning. It needs more work. I need to add an option to limit it to a single gif, right now, if you do a grid of these frames, the results are pretty good, but it'll also pump out a bunch of garbage gifs for the rest of the images. Not good.
So. It works, Well. just keep that in mind.
\o/
Any workarounds for when this misses the central image especially when working on non-anime checkpoints?
Hey this is really cool! I do think there needs to be a more reliable way to remove the white bar or edit them out of your training pictures
After using this a bit it is insane how dedicated it is to keeping that white bar!
@Galtherok349 Yep. There aren't many sources for that clear same character before during and after, I spliced a lot of works from like three different artists together, and all of them used that white bar heavily.
If I do the training stuff again, I might run an AI uncensor over the data set and retrain from scratch.
@emberbyte I'd be super grateful! Sadly I know almost nothing about training Lora's but besides the line this one is awesome, the images all came out clear and the character was consistent in each frame.
In the mean time I might try to find a program for removing the censor on the outputs
What animation addon do you use
mizumizuni!!!!
not sure why but it doesn't work for me.