Version 1.2 Update: Higher quality and expanded dataset. It can now do cumshots where the woman keeps sucking and you can see the cum dripping out of her mouth. The I2V version can handle images where a penis does not start out in the image.
Note that all of the previews were created using the lightx2v (self forcing) lora. You may get better movements without it, but I didn't do much testing without it. The workflows I used can be downloaded by getting the "Training Data".
T2V tips: It actually seems to still maintain motion very well at 0.7 strength, and you'll get more facial variety that way. I seem to get best results at 0.95 strength for whatever reason. You can add in a penis lora to help it out.
I2V tips: If the penis does not exist in the starting image, you probably want to add a penis lora to your workflow so it has a better shape. You can trigger it with something like:
A penis appears from the bottom of the frame centered at the bottom and pointed straight up.If you want cum to appear:
White cum shoots out of the penis.This was trained on 4 different angles, here is how to trigger them:
A woman is lying on her stomach between the legs of the viewer and performing oral sex on a man. Her head moves up and down as she sucks the penis.An overhead view of a woman kneeling between the legs of the viewer and performing oral sex on a man. She moves her head back and forth as she sucks the penis.A woman is leaning over a man positioned in between the legs of the viewer and performing oral sex on a man. Her head moves up and down as she sucks the penis.A woman is kneeling in between the legs of the viewer and performing oral sex on a man. Her head moves up and down as she sucks the penis.Description
FAQ
Comments (29)
thank you for posting your workflow too. with the workflow i had been using the lora was not working well but with yours actually works.
where did he posted the workflow??
@eroSennin3487 read the description. it's in the training data download.
i2v 1.1 with this lora I alsways get some sort of pixelation with the first frames and then the face changes and is totally different from the starting frame
No idea why that would be. I haven't messed with I2V so I don't have any advice, but it seems like people have used I2V successfully based on videos uploaded to the gallery, so I wonder if it's something in your workflow causing issues.
@dtwr434 I use wan in Pinokio (without comfy) , so I have more or less no workflow
I don't know anything about Pinokio, but I'll copy and paste a message I got from someone else on another one of my loras:
"It might have been the version of Wan2GP on pinokio. As of the March 13 update which specifically addresses loading LoRAs, I'm getting great results now. Using CFG 4, shift scale 6, 20 steps. On my 4080, can get up to 150 steps. Fast and coherent. Thanks!"
So maybe you need an update? Otherwise, I'm sure it's not what you want to hear, but I'd say switch to comfy, as that's pretty much the best supported software and will always give the best results.
@dtwr434 yes WanGp it is I use, Pinokio is a wrapper. I made serveral tests with same settings and same starting frame but different loras. Some other loras give me the same issue. The nsfw lora for exaple gives me far better results, even for bj's
@losveratos102486 Alright, I finally got around to trying I2V. I used this workflow and it worked great for me: https://civitai.com/models/1297230/wan-video-i2v-upscaling-and-frame-interpolation. The faces remained consistent and the movement worked well. I'd maybe have to play around for a while to figure out the absolute best settings for quality and things, but overall I'm impressed with the output in my first few attempts.
So, I don't think there's anything fundamentally wrong with the lora. If you're using the exact same settings for every lora, that could be the issue, as at least in terms of strength, it probably makes sense to adjust that per lora. Not every lora works best with the same strength value. I used 0.9 for this one as a starting point, though I don't know what is ideal from my quick testing.
I also wonder, are you using a realistic photograph to start with, or is it anime? If it's anime, that might explain why the NSFW one works better, as it seems to have a more anime focus.
@dtwr434 yes I used a good foto, several even and also testet with lower strengh settings. it also seems to me that the face always morphs into the same face, has the lora perhaps been overtrained?
@dtwr434 I found out RIFLEx is the fault, but it seems it doen't effect every lora
Ah, interesting. Glad you were able to figure it out.
@losveratos102486 I have the same issue, are you sure is Riflex? I deactivated it and still get a realistic image when i set an anime one, what settings are you using other than that
for i2v, i find that the girls always seem to bite a chunk out of the dick and don't really follow the motion. i might have to try another workflow
I had success with this one, though I didn't test it extensively: https://civitai.com/models/1297230/wan-video-i2v-upscaling-and-frame-interpolation
@dtwr434 thank you, i'll try it out
@dtwr434 it was def the workflow, loras didn't work well with it. unrelated question: i've noticed that the faces with this lora kinda all resemble each other, even if i change facial features in the prompt, is that somthing that can be addressed at some point (assuming you're going to update it)
@crombobular I wouldn't know how to. There's already a decent amount of variety in the dataset. Probably character loras or I2V could help with that.
Can you post a workflow of how you make women bite off the dick?
Asking for a friend.
What's the difference between yours and aipinups69 one? Looks like yours is dated later but did you train yours more then him or on different positions?
v1 of this one is earlier than the other one, and v1.1 is a refinement. But this one is a POV focus, while it looks like the other one has more of a side view focus. It never hurts to have multiple loras for the same concept anyway, since it will lead to more variety.
can you make a 13.b fun version?
This generally works well, and doesn't have the fade-in/out and other artefacts that I experienced with the other main Wan BJ LoRA, but I still have issues with the morphing of the penis end. I guess you either have to prompt your girl to never pull all the way out or just pick out the 25% or so of generations that don't show the head going all morphed.
do u think ur workflow will work on 8g vram?
you Model should be a GGUF 4-5.5 go. Like wan q3
You know how AI can't make a wine glass that is full? I feel like it's the same with blowjobs.
have you recently tried generating a full wine glass?
zbzbzbbb yes. Up until very recently it was not possible. A wine glass is only supposed to be filled halfway, so to a computer that is "full". If you don't believe me, google it.
DootyBoot Yes, I am aware of it - surprised myself when generating a full glass a while ago.
I do get your point though, technology will eventually let us adjust the depth more easily ;)
Details
Files
pov_blowjob_v1.1.safetensors
Mirrors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors
1_pov_blowjob.safetensors
22_pov_blowjob.safetensors
106_pov_blowjob.safetensors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors
pov_blowjob_v1.1.safetensors