Deepthroat Blowjob LoRa
Trained on a dataset of 78 clips with 224x224 resolution, 1sec to 3sec
“pov” and “sideview” camera angles (pov seems to be working better now)
22 different source videos.
also i just finished preparing a new dataset with different camera angle, stay tuned for when its ready 😎
* Added v1.2 with "normalized" dataset, i mean all clips split into 3s - 1s clips, and "re-captioned" all, learning rate of 2, also "video_clip_mode=multiple_overlapping" in hunyuan_video.toml this time)
The workflow im using is:
Hunyuan💥AllInOne ▪ Fast ▪ T2V | V2V | I2V* | Upscale ✔️AIO | Advanced Tea☕1.4 from @LatentDream
Examples:
a sideview of a woman giving a deep throat blowjobsideview of a woman kneeling in front of a man deepthroating his penis, she sucks deepthroat style, woman sucks a big cock, she swallows the whole penis, deepthroat all the way, she looks up at viewer, balls deep blowjob, deepthroat, lips on hipsa sideview of a woman swallowing a penis in deepthroat, the woman is kneeling over the penis, she swallows the whole dick touching the testicles with her chin, she sucks it from tip to base, the man grabs her by the hair, side_d3pthr0AtPOV view of a woman kneeling in front of a man deepthroating his penis, she sucks deepthroat style, woman sucks a big cock, she swallows the whole penis, deepthroat all the way, she looks up at viewer, balls deep blowjob, deepthroat, lips on hips, pov_d3pthr0AtHave fun! 😜
***** This LoRA is shared without warranties and with the condition that it is used in a lawful and responsible way. I do not support or take responsibility for illegal, harmful, or harassing uses. By downloading or using it, you accept that you are solely responsible for how it is used. *****
Description
Trained on a dataset of 46 clips with 224x224 resolution,
“pov” and “sideview” camera angles,
22 different source videos.
4018 steps and 14 epochs.
FAQ
Comments (11)
Oh a side view very interesting, I'll have to try this one out.
What frame selection method did you use for your videos? How long were the clips?
Appreciate any suggestions. My sodomy lora works but it gets choppy and suffers from lack of motion so I want to try to modify my data and methods.
Thanks for the awesome lora.
the fps are all 24, as for the length i actually mixed 2s, 4s even 8s clips, the diffusion-pipe handles it nicely, next one i will try 512x512 resolution, if i dont get OOM
@K3NK - do you mean there is no option in DP to adjust the frame extraction method?
I'm talking about this:
• frame_extraction_method (str)
Description: How frames are extracted (start, chunk, sliding-window, uniform).
Choices: ["head", "chunk", "slide", "uniform"]
Default: "head"
• frame_stride (int)
Description: Stride used for slide-based extraction.
Range: 1–100
Default: 10
• frame_sample (int)
Description: Number of samples used in uniform extraction.
Range: 1–20
Default: 4
from here:
https://github.com/zsxkib/cog-comfyui-hunyuan-video
@Gongoloid Diffusion pipe gives you this options:
# How to extract video clips for training from a single input video file.
# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
# number of frames for that bucket.
# single_beginning: one clip starting at the beginning of the video
# single_middle: one clip from the middle of the video (cutting off the start and end equally)
# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
# default is single_middle
video_clip_mode = 'single_middle'
im using single_middle
@K3NK - word, thanks for your answer.
Cheers!
I've got another hardcore lora brewing...
@Gongoloid i think this is the issue im having with the POV perspective, i have single_middle so it might be that, this setting is cutting out the last part of some clips i guess, being that part of the clip where she goes deep..., i might try re-retraining with new settings
@K3NK - I tried to run my first video dataset last night instead of frames, and just couldn't troubleshoot my errors. But images are training fine, so I went through my 16 clips that I had already extracted into frames, and I selected 20 contiguous frames from each folder, so I have 320 frames. Running through them it looks like a solid idea IMO.
I'll let you know how it turns out.
Should i try training this dataset only using keywords instead of describing the scene?
Does anyone knows if this is better for action/concept loras?
At the moment im further training more the model, to get to 10k steps, idk if its good or not, ill check when i come home from work. I wasnt able to reduce the learning rate when resuming the checkpoint, diffusion-pipe is at 5 i wanted to get down to 2 but i need to modify train.py script and i dont want to mess It up 😅,
I asked gpt but It stopped my chat coz yesterday It generated a random image at some point (i asked It to generate a city ambient sound and It generated an image!!! With a play button that i pressed like an idiot 🤣) that noone asked for, then told me that It cant generate sounds... now the chat conversation is blocked (cant continue with gpt3, coz he previusly generated that image that noone asked for... Insane 👿), GPT4 Will try hard to spend all your free GPT4o tokens doing bs... Better be aware of that.
someone more qualified will surely answer you but i would suggest you to embed to workflow with the video so that we can get exact results like yours using ur settings most of the creators here are doing it so please try. Great work with the lora regardless
Using keywords is always better.
Yes, the official repo says to use full paragraphs (and it was trained this way), but this a negative for diffusion based models. This is because the meaning of the tokens is spread out and having "walking towards", "moving towards", "approaching" are going to have wildly different results because they relied on an LLMs description, not yours.
Less words = less token variety = more consistent results = better. CogVideo was trained the same way and it drives me up a wall.