👄 Blowjob Concept LoRA (Video | WAN2.2 5B)
Hey again!
This is my second video LoRA, trained for blowjob-style animations using WAN2.2 5B. Like before, I used Ostris’ AI Toolkit and followed the same official tutorial. The process was simple and smooth.
It is actualy more deepthroat then blowjob I forgot to change the trigger word.
🔧 Training Details:
Model Target: WAN2.2 5B
Dataset: AI-generated 3-second clips (72 frames) using WAN2.1, focused on oral scenes
Can generate both side view and POV angles
Resolution: Trained at 512 this time.
Training Steps: 2750 total
Training Time: ~4 hours on an RTX 5090
VRAM Usage: Peaked at around 14GB
Workflow: Default ComfyUI video training workflow, with LoRA loader
Trigger Word:
bl0wj0b
💡 Notes:
While WAN2.2 5B isn’t as high-quality as 14B, it’s quick, lightweight, and gets the job done.
There's still room for better datasets and fine-tuning to enhance facial and mouth motion accuracy.
Hope this helps others trying to make animation LoRAs!
Feel free to test it, give feedback, or ask questions 🎬💦
Description
FAQ
Comments (30)
Does this lora only work with 5B? The details on the right side of the page says 14B.
Civitai did not make a wan 5b category.
if is 480p, why your videos are 1500x2500 48fps?
Plus, if I download them (mp4) Windows cant play them. Neither Avidemux can opne them... What coded should be?
Is the workflow attached? my 5b results were never good i wonder how you did it...
model is good but tends to create long fake eyelashes :/
It was trained on generated wan2.1 videos with cosplay
iPiKo yeah so it kinda needs to blur faces where possible to prevent lora affecting generations this way
So what are the advantages and disadvantages of using the 5B model vs the 14B models?
Easier to train and faster inference speed without any hacks.
The quality of the video I think it is not bad I still need to try the fp32 version to see.
It is bad at prompting and doing what you tell him to do, it also do not have a level of understanding of animation like 14B do but with LoRA this is not a problem.
It's "okay". I guess it's useful for people with low (8GB) VRAM; but even then, quantized 14B models would probably be better. If you have the VRAM, use 14B models every time.
I'm trying to use your LORA with WAN 2.2 5B 720p on WaveSpeed (a online service) but no movement at all. Is like the LORA did nothing. I tried weight at level 1 and 2, nothing... the strange fact is others LORA works, even LORA for WAN 2.1... I cant understand... are you sure the prompt is simply bl0wj0b, like in your examples, and the LORA will do the job?
The prompt is not only bl0wj0b I think you don't even need it, you need to describe the scene download the video and drop it into comfyui to see the prompt.
iPiKo Hello, meanwhile I noticed there is a (new?) category on Civit AI. Is named WAN 2.2 ti2v 5B. "TI" I presume should include both "t2v" and "i2v." May be is better describing your LORA?
settima_ai all my lora are i2v
iPiKo Perfect now! Thanks! : )
iPiKo Dragging mp4 files into comfy doesn't work to show the workflow, never works for me. Only webm files do.
epic cool
Using Wangp, but i put the lora in, set it to 1.0 but nothing happens, is there another lora needed nsfw?
Not sure I never used wangp but you need a picture to generate.
iPiKo OK. Got it, doesnt need the start image to contain everythig to work. it works now.
I need .lset file as well to use this .safetensors in wangp. Any suggestions?
weird face after gen, the character when wrong but the movement is ok, hope you fix the lora soon
hey! which software are you using for finetuning? Somehow I can't find anything about the 5B finetunung ....
I did no finetunning I have linked the tool and tutorial for the training.