The premise here, is that I want a slider for how deep it goes. V3 almost does that.
Each of the individual stages, when used with their tags, can are deeper the higher the pressence. Ie. (stage five, fellatio, deepthroat, ballsdeep:1.6) is deeper than (stage five, fellatio, deepthroat, ballsdeep:1.1), I'm going to retrain again, and see if I can get a single keyword to work as that slider.
staged deepthroat and fellatio
triggers
stage zero: imminent fellatio
stage one: imminent fellatio
stage two: fellatio
stage three: fellatio
stage four: fellatio, deepthroat
stage five: fellatio, deepthroat, ballsdeep
stage6: after fellatio
trained on both the stage X words, and the tags; works better with the tags
if you want to turn them into gifs, like the sample, check out my animated gif helper scripts, grab the prompts_from_file_to_gif script.
Description
Third version! The second isn't even worth mentioning.
This is a lot easier to control than the first version. Still not perfect though.
I realized that I'd borked my tags, so this one is all wd tagger. I'll probably go back grab the booru tags, add any new ones, and then retrain.
FAQ
Comments (11)
TIP: use this with style loras, it's pretty easy to make coherent animations with V3 if you use a style lora with it.
i use 3 different kind of model and still not work very well
@ShonanFM Did you try the prompts in the samples?
I trained it on NAI, and the samples use AOM3; so it should work well on any model mixed with NAI.
And if you look at the gif, I have each frame labeled with tags that should help you get close to that same affect.
Also! I just finished training a V4, haven't tested it yet though, so depending on how it turned out, there might be a better version up soon.
V1 and V2 I had borked my captions; V3 has all WD tags, and V4, has both booru tags and WD tags, so hopefully it turned out well.
V3 shouldn't be too hard to use though, unless you're trying it on purely realistic mixes, then it might not be great.
@emberbyte I try the nai.final model and it works.
But the character is change and unstable, and when i try to test the other frame, the huge frame change work is a mess.
Can you add a blank UI to add this extension as (action image)+(user promp) in the script?
You can reference the urthor BDinc's lora timelapes. according to his scrips:‘{character_prompt}’.
@ShonanFM This isn't a character or style model, I made this to be combined with those.
I have others that affect style and character a lot stronger, but that wasn't the goal with this one. I wanted it to be a bit more free.
The goal with this is have staged step by step depth for deepthroat, which it does.
In order to get a consistent character, you need to either add that information to your prompts or use a character or style lora, a good way for prompts to do that is prompt editing celebrities, ie
[cameron diaz:catherine zeta jones:demi moore], green hair, red eyes, skinny
If you take the prompts in my prompt file and add that to the end of each line; takes seconds using find replace in a text editor, you can use them in the prompts to gif script on my animation helper page.
As for writing an extension that does that helps you do that?
Sure. If I remember :), I'll get on that after work tomorrow.
I thought of a decent way to make it generic enough to be valuable. I can start including template prompt files with my models, and then have the extension simply do a find replace on the templates with the input character information.
For that. You would enter your character description, and any lora you have for them into a box, and upload the prompt file, or enter the prompts into a different box, and then poof, generate gif.
That's basically what the prompts to gif script on my animation page already does, just without the find replace option.
Could you make a lora of deepthroat with head lock by legs?
Umm i did create around hundred of images to test the staging, it turn out the size of the D affect how deep the staging is, especially stage 3 being exact same with stage 2, and stage 5 like stage 4, while for the small D stage 5 and 4 are a bit same but still balldeep, and stage 2 deeper than stage 3 which also the same like stage 1, I tested them in XYZ plot with prompt S/R with these weight which combined with CFG, hi-Res, and seed : "(stage zero: imminent fellatio, spoken question mark, happy:1.2)","(stage one: imminent fellatio, spoken question mark, saliva trail, happy:1.4)","(stage two: fellatio, raised eyebrows:1.4)","(stage three: fellatio, spoken question mark, expressionless:1.4)","(stage four: fellatio, deepthroat, spoken question mark:1.5), (angry:1.1)","(stage five: fellatio, deepthroat, ballsdeep, spoken exclamation mark:1.6), (rolling eyes:1.2)","(stage6: after fellatio, exhausted, facial, cum in mouth:1.4)" ah and also it seem it worked best with H model checkpoint like hassaku.
had to crank it down to 0.4 to keep the character lora from being a bit fried, but in general it worked better than the 'pov fellatio' lora that's on here :)
Why are there no heads in any of my images，just bodies！！！
The stages seem to make very little difference in ym experience.
Pretty well done, but it's actually almost impossible to reach proper "balls deep"