Man! What can I say?
This is absolutely fantastic! You just have to try it to truly see the difference compared to the others.
V2 has released without face locking and more different action on Oral performance!
no face locking, improved asian face vairance
more eye contact to camera
multiple different way to lick and suck the penis, need try out
easy to combine with face lora with the node from my friend Fok:
There is workflow metadata in the videos! Right click and choose "save video as", and NOT "save videoframe as" which is indeed empty
Support me:
# Support Me!
If you think it's good, remember to support me!
Here are my contact details:
- Civitai: [ttplanet](https://civarchive.com/user/ttplanet)
- Discord: ttplanet
- X: @ttplanet
Contact me if you want your customized lora!
Highlighted Features
Full-Speed Movement: No sluggish actions like most LoRA or video models.
Realistic Style: Delivers a genuine POV experience.
Customizable Characters and Backgrounds: Easily replace or combine them as you like.
Super Realistic Expressions: Natural-looking faces with eyes properly focused on the camera.
High Robustness: Almost a 100% success rate in generating videos!
it also works for man characters....
Areas for Improvement
Facial Variability: Currently, faces are too rigid, especially for Asian features. Will address this in the next update.
Proportionality: Some generated elements appear oversized (e.g., certain body parts). Planning to retrain with more appropriately scaled samples. If you can provide relevant material, it would be greatly appreciated!
Description
FAQ
Comments (26)
Tried it out and works nicely! But you are right. The face is pretty stiff. I have the feeling that loras that are trained for one position and view only have trouble to generalize people. So a blowjob loras with many different angles have a easier time to generate different kind of people because there is much more context to choose from. And then you can still focus on one view with your prompt !
the turth is I have over 15 different people’s face! multiple clip for each one. this may due to the reason you mentioned. I will need to add more clips to it! just like I said, I think 200 clips is a proper number to get the job done.
@ttplanet maybe having the same person in multiple clips also burns in their appearance? But cool that you keep wowrking on it. Looking forward to your next version!
@WhatTheGuy I got a addtional 50 person added with much more different action, but I need to tag them, what a mess job.
@WhatTheGuy again, the apperance look like none of my training set....
@ttplanet yes a mess job, but that's the difference between a lora that's just good and one that's awesome ;)
豬佬,請問這個也是透過Gui_for_musubi-tuner訓練而來的嗎?這種動作連貫的lora不知道該用什麼樣的素材訓練...
使用video训练的，我提前准备了很多适配的素材，我正在增加素材库，重新训练
@ttplanet 謝豬佬回覆,Gui_for_musubi-tuner也有支援使用video素材?還是說使用別的工具訓練的?diffusion-pipe也沒有Gui版本,用起來各種不方便
@ioritree 我分叉了diffusion-pipe並製作了一個簡單的gui，除了我添加了一個docker鏡像，如果你想測試它，請點擊存儲庫鏈接，目前如果你使用docker，它是100％功能的，我仍然會調整一些東西無需docker即可運行。alisson-anjos/diffusion-pipe-ui: A pipeline parallel training script for diffusion models.
@alissonerdx thanks
Ah I saw you have the same problem as me! At the end of all your clips there is a short ghosting of the image. Mostly visible in your first example. I just reduced the video to 61 frames instead of 72. Now my videos don't end with that annoying ghosting anymore ;)
because the my max train fram is 65
weird. after over 2000 tests i ended up choosing 65 as a frame amout that give me best results. so looks like is a training limit. uhm interesting
@LatentDream hah seems around 65 is some cut in the training. Like buckets, but not for image size but frame size
I've noticed for some lora it's possible to get rid of the ghost frames by reducing the weight, but it's difficult because it can be too low and the lora stops working. I haven't tested this one yet, but for the titty drop one it's the case. Maybe it's because it's too new, or an issue with the training. Has anyone tested musubi, from tencent promoted kohya-ss/musubi-tuner ?
@ttplanet does this mean that for training it's better to use 100-125 frames if the goal is to generate 100-125 frames videos?
@bbeggs234 we got to find a solution, else loras will become messy, as you can never know if someone trained it long enough. Then you have a ton of lora with frame glitches after 65 or so frames. Say you want to use a lora on GPU and generate say 10 seconds or something like that and it's generating multiple ghosting parts. That's bad in long term. Hopefully that's only an issue with a certain way of training that we can guide people to avoid.
@lost_moon I knew the reason and this is my choice to balance the resolution and frames based on my current hardware. But I do have technical way to reduce it as you can see V2 perform much better due to it.
Can you include your prompt?
can you share your workflow?
inside the video
@ttplanet I have tried to save the video you shared and dragged it into ComfyUI, but it didn't work, can you tell me why?
Getting great results, thanks, what kind of criteria did you use to select your training videos and how did you caption them? I want to emulate it for other motions