A lora that attempt to capture a sort of gratuitous male finishing, h-style. It is intended to work with all sorts of positions/actions.
V0.2:
Probably less vomiting?
Probably more knowledge of positions? All samples were generated with no additional loras - this is in some sense a (mid) general nsfw model I suppose.
Probably worse motion for new positions (due to expanding dataset to be not only goat creators)
On the whole probably? maybe? better. Let me know if you think otherwise though.
Some neat functionality:
It knows different parts of the male appendage, so you can describe actions along the "shaft" or "tip" for certain things (I've posted one for grabbing the appendage).
It also generally seems to understand the moment of release, so you can specify actions with time-relation to the moment (before, after, during).
V0.1:
At the moment it doesn't exactly seem to know where the end of the male appendage is 100%. It also sometimes adds fluids in places it shouldn't. I think I need to do another pass at the labeling, but for the moment it at least works ok.
Fun side effect: It knows how to do zoom ins. See the last example for a prompt to do that. Let me know how it works for you. I recommend 0.75 strength, may be a bit over tuned for higher strengths.
Importantly, see the trigger.
Description
FAQ
Comments (30)
mind sharing your workflow?
Sure! For data collection I just manually download some vids that I have in mind. I normally don't start with a concept but start with some quality video, and figure out what I could make from the video. That way I always have high quality data. I trim clips of slightly longer than 5s using the windows photos app. Then I use shutter encoder to trim every clip to 81 frames, and conform to 16 fps, as I understand wan 2.2 is trained to produce 16 fps video.
For labeling, I just do it manually in natural text describing what is happening in the scene. It takes a long time, but I'm not aware of a better way. I use ostris/ai-toolkit to actually do the training, which handles resizing (otherwise I would do this earlier with shutter encoder). I train at 300 resolution, as from my experiments that's the highest I can go without running out of memory. If I had more memory I would go higher, but I refuse to use any cloud services for this type of stuff. I think my settings for this lora were the default outside of the resolution, and that I trained for 4675 steps.
Let me know if you have any other questions!
@bobbysmithy55555426 What is vram needed for training and how long is the training 🤔?
I have 32gb vram. I know others get good results training at 250 resolution on 3090s, so that's possible if you have 24gb vram.
And they take like 20 hours for me to train lol. I just leave it running for a whole day pretty much.
Thank you so much!
Can you recreate it in this way? ttps://78.gigafile.nu/1220-dfbcb6e2fa5fbcb731366eb9ba4db4ab9
Very often, the mouth vomits fluids at the same time. Any way to prevent that?
Yeah its a bit of an issue. I find using it at .75 strength instead of higher helps, but I think I need to do another pass at labeling and re train to fix. Apologies
Try to reduce the weight of the high LoRA or set it to 0. In my case, it worked out quit well.
If I use it for 2.1,which lora I can download it?
Sorry only trained a 2.2 i2v. I mostly do this for personal interest, and can't be bothered to devote 20 hrs of training time to any other model.
@bobbysmithy55555426 i wanna ask, if lora trained by 2.2 so i can't use the lora in model 2.1?
@Gooddemon I've tried it with 2.1 and it does generate but my results seemed more experimental than polished. Not sure if I'm using the right keywords and settings or if it's really meant for 2.2. I'm moving to 2.2 so I'll try it there and see if it's better
Edit: I use it with 2.2 and the results are vastly superior. I'm still playing with the strengths because sometimes cum comes out of strange places but it's still better than 2.1
Cannot make the motion stop and cum like the example. Did you cut out the close-up part of the video? Because those words are in the prompts but not seemed to have effect?
Anyway I hope you improve this lora so it works consistently, it's so peak if it actually works like the examples :)
Are you using other loras at the same time? I think for me to get it to stop like the example I was using only this lora (at .75 strength), or this lora with my "action from behind" one at very low strength (like .25). IIRC using the general nsfw, or my other lora at higher strength would both trample on that effect.
For the examples using the close up text but not doing that, I think I may have copied the wrong prompts? Civitai doesn't properly copy the metadata when I upload so I just go back and do my best, but possible I used an adjacent example here. I know I was also trying to see if the lora had learned to cut to an entirely new angle (not just zoom in) which was ambitious, and unsurprisingly doesn't work, so that could also be what's going on here.
Anyways hope you can get it to work, let me know if that doesn't help
Is there a realistic version of this or do you intend to make one?
Sorry I only make anime stuff for now, no realistic version coming
It seems that the quality of the liquid is relatively poor; the pure white color lacks texture, but it is still excellent.
Yeah I imagine I could get better results if I had more vram at my disposal (to train at higher resolution), but alas
@bobbysmithy55555426 May I ask how much material you used and about your equipment situation? I also want to give it a try. If the effect of this lora is better, it will really be very useful.
Sure, I've got a 5090 and this was trained on 44 clips. I put some more detail on my workflow in an earlier comment on this post as well.
For what it's worth, I've found that I get somewhat more realistic looking cum effects by specifying "sticky translucent clear goo" in positive prompt and "glowing white liquid" in negative.
*PEAK* You totally deserve 3 burgers for this, just saying. 🍔🍔🍔
And keep up the good work for us, bobby❤️
Thanks, peak is the best compliment I could hope for
if you can nipple tweak lora may I ask you a favor?
Regardless of which model I switch to, the resulting liquid always appears as an unattractive, paint-like white without any gloss. I believe the liquid material still requires improvement.
Yeah I think this concept would probably benefit from more vram, as the effect takes up a relatively small portion of the frame and so suffers more from being downscaled to a lower resolution than some full screen concept would.
But I'm just a guy with a 5090, and I'm not going to pay for cloud services to train cumshot videos lol, so this is probably the best you'll get from me.
I don't always get good fluid out of this, but there are cases where it helps quite a bit with motion.
excellent job,due
I'll try it out. Your first version was pretty solid
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.