My first attempt at training a lora. It turned out well enough to share, I think. Trained on 20 5-second clips at 512x512 resolution. The intention is for it to be a lora with more of a focus on thick lips working the head of a penis. Trained primarily on pov shots, with some three-quarters and side shots. If you have tips for motion lora training, I'd love to know!
It's not the most consistent at forming a good-looking glans when there isn't one in the initial image, so I recommend pairing it with the low noise side of a lora good at that, such as Ultimate Deepthroat. It seems to work well enough without the trigger word.
All showcase videos were made using this lora on high and low at 0.8-1.0 strength, but I recommend testing different strengths. My preferred generation method is to use multiple loras at lower strengths, but I wanted to clearly demonstrate this one in the previews.
Generations in the showcase use 4 steps high, 4 steps low. They also use the following:
Workflow: Smooth Workflow v2.0 - https://civarchive.com/models/1847730?modelVersionId=2264611
Checkpoint: Smoothmix i2v 2.0 - https://civarchive.com/models/1995784?modelVersionId=2513182
High Noise lightx2v 1022 - https://civarchive.com/models/1585622?modelVersionId=2337890
Low Noise lightx2v (1.5 strength) - https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Lightx2v/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16.safetensors
Description
Initial low version
FAQ
Comments (20)
Oooooh. Looks good.
Downloading to test. We can never have enough good blowie LoRAs. ;)
excellent work! and it doesn't change the face which is great... looking forward to your future work
This is exceptional, good work!
DAMN good!
Great job on your first LoRA. Congrats! I did some testing with a GGUF base model and you absolutely need another LoRA for the glans when the penis is already in her mouth. else it'll look like a sausage. I chose Oral Insertion 0.5 on H and L and with this combo I'm blown away how great this LoRA is. It works fine when the glans is visible in the base image though. But since I'm lazy I kept the LoRA-Combo for all my runs.
Thanks so much, Alberist. 🤜🏻🤛🏻
Another fine addition to my evergrowing BJ/Deepthroat LoRA-Stack.
It's interesting that you say this because in my experience "already in her mouth" just basically never looks very good with any combination of LoRAs I try, at least nowhere near as good as if you start with the full context in the base image.
@StrangeLimina I find the are a few loras that produce good looking glanses if you use their low noise lora on higher strengths. Though you definitely get more detail and control if you just have it in the initial image.
@StrangeLimina For me Oral Insertion makes the glans look good most of the time.
@fatberg_slim @Alberist Interesting I'll give that a try, thanks for the tip (heh).
Very well done, especially for a first Lora!!
Nice job identifying a need. Most LoRA's don't seem to have enough face diversity to capture this. Looking forward to trying this one.
I'm testing. So far all my gens have been very different than everyone else -- tons of head bobbing, like 6 bobs in a 7 second clip. Been tweaking strength and it seems to need a little help LoRA for the glans as Fatberg was saying. (I went with experimental NSFW).
Yeah. Oral insertion, Ultimate Deepthroat, and Big Cock Oral should all be reasonable ways to improve penis shape, depending on your preference.
@Alberist Your LoRA takes "she is bobbing her head multiple times" and runs with it! Maybe I need to adjust the shift to get it to work. But taking that out, and I'm getting the real thing. Again, great job on this LoRA.
Top notch bj lora this, 10/10. Good cohesiveness, minimal facial morphing. Will certainly be using it lots in future, combined with dreamley and the penis glans was just fine for me.
Hey there^^, can i ask what setup you used to train this model? (GPU/RAM/CPU).
I rented an H200 for a little over a day and trained in the cloud.
@Alberist ah ok, i was trying to figure out if i could somehow do it with a 5090 but until now no good resaults sadly
@TheLuckyBunny I wouldn't give up! @TwoMoreLurker put out this guide on lora training on a 3090, and their results still come out good. https://civitai.com/articles/18985/wan-22-local-lora-training-guide-windowslinux
I do plan on trying to train with my own 5090 in the future, but I'm usually running generations at any given time.
@Alberist oh thats awesome, thanks! will look into that the next few days. if you are ever interested in doing a cooperation (maybe trying to develop new character or animation models or working on workflows) i am always open for it, currently i am mainly working on a Discord where i share stuff that i make and where people can request stuff.