My first attempt training a video lora
Trained on Wan 2.1 i2v 720p
Example prompt
``her breasts start to grow smoothly and quickly, straining against the shirt, and causing big cleavage``
I used same prompt for all the example videos
Configurations
Steps: 20
Cfg: 6.5
Sampler: uni_pc
Scheduler: Simple
Haven't tested for t2v
You can try this trigger word 'BR3A5T3XP' i2v doesn't need the trigger word but I'm not sure about t2v
Description
FAQ
Comments (82)
You beat me to it! Haha, this looks great, can't wait to try it. Having trouble getting training set up but have my dataset all labeled. Curious how many videos or images you used?
8 videos, between 3 to 5 seconds length, I'm new to it too, good luck for your training 🥰
@CaptainWebSpinner Very nice! Did you do locally WSL or runpod? I keep getting OOM with 24 gigs of memory so may need runpod.....
@katsumend I did it on TensorArt website 😅 my pc can't even generate an image
@CaptainWebSpinner How do you find these training videos? Thank you.
@ysc3839 There are alot on YouTube and civit, you create your own too I have written an article about that you can check it out. And Thanks for the buzz ❤️
@setoror Thanks for all the buzz, it's really motivating ❤️🥰
!!!!!!!!!❤️🥰!!!!!!!!!!!!
Finally its here for Wan!!!
Was anyone able to download the "SuddenOutfitChange_V03" Lora from yesterday? I received the model and now need the Trigger Word. Comment or PM me if you remember the trigger word and 480p or 720p model. I missed that one. PM me a download and I'll add it to the nsfw archives list https://civitai.com/articles/13990/join-the-nsfw-ai-archives
BTW - Sick lora, even trained on 720p, which is incredible!
Trigger: then a sudden outfit change into
Example Prompt: A girl is posing, then a sudden outfit change into a traditional Catholic black and White cuff Nun Habit including a matching black veil that symbolizes purity and dedication, her large breasts are bouncing up and down, she is smiling
@null_name MVP - Thanks! I'll test it out to see if it work then uploading to the Archives
@null_name I can confirm this to be correct. HASH: 50DAC7703, creator: ai_build_art
Used I2V-14B-720P as base.
@MagicalErotica @null_name Awesome thanks guys, work perfectly
https://civitai.com/images/73430695
Google sheet updated:
https://docs.google.com/spreadsheets/d/14P_gbpCtr0HZ25KAGa9M_1_5cEO6GJGX_m3778GStlk/edit?usp=sharing
Hi ! Is it possible to enlarge breasts under clothing without showing cleavage ?
I don't think it's possible, the dataset didn't have any such videos, basically it's trained on clevage
@CaptainWebSpinner Ok. Thanks for your quick reply. Another thing, is there a way to play with the breast size by changing the parameters?
That's my shit right here. Thank you so much. Take my buzz
Hi! I'm attempting to learn I2V training, did you usee Musubi or Diffusion-Pipe? And were you able o train locally or did you use a cloud service like Runpod?
I used TensorArt
@CaptainWebSpinner nice works pretty well , curious about training settings as well
it seems like you used
Network Dim : 64 ? Network Alpha - 32 ?
What about settings ? just default ones on TensorArt ? ty)
@blo01 Yes, network din 64 and network alpha 32, Target frame 13, Frame sample 5, I used default settings for other parameters, Also I blurred the faces and added "censored face, blurred face" in the caption
@CaptainWebSpinner oh that's interesting , might try blurring objects that are not trained for to avoid added noise they gives
@CaptainWebSpinner I've been working on a similar lora using my own training set and the civitai wan lora trainer. This is my first wan lora with motion using videos for the data set and I'm not having any luck. In your dataset captions, do you describe the motion ("her breasts are growing", etc.) or do you only describe the rest of the scene? There is very little information online about captions for wan motion lora. Also, how many videos, repeats, and epochs did you use?
@darkroast175696 Civit's trainer doesn't work for me too and No I didn't mention that her breasts grows in the caption I just described rest of the scene and used a unique trigger word (br345t3xp)
@CaptainWebSpinner That's very frustrating, thanks for the info. I used civitai's trainer for a character lora and it's working very well but that was using images for the dataset, no videos. So maybe their video-based training is broken? It's pretty costly in buzz to experiment so I probably won't try again until someone says they can verify it's fixed. Thanks for the quick response and also for the different Loras - I've had great success with all your models so far.
@darkroast175696 Yes, it works works for character, clothing and Style loras but doesn't work for action loras, I will be experimenting now and then, will inform you if it starts working or if find a way to get it to work
it worked once and now every image just turns black after like 6 frames?
I didn't have that issue with this lora yet, it usually happens with me sometimes while using wan, try reducing cfg it could be the issue
if i run it with another lora it works fine or i have to turn the cfg up to 6.5 its something in my preffered work flow and not sure what. i consistent results when i used it on a basic set up
Look forward to trying it out, thanks! it's named with code though, you should give it a human-friendly name in the next version.
Lads, this is EXACTLY what happens when your wife/gf gets pregnant. It's not 'exactly' this fast but close (roughly x0.25 or faster!)
PRO TIP. If you watch her real closely, you can see them baby-nozzles increase in real-time. Don't blink.
Excellent Work. Thank You.
I'm kind of terrified that millions of boys are growing up without even knowing what a breast feels like, let alone experiencing real-life expansion. It's not something that can be substituted for. Well, afterwards. After she turns into a hellbeast and starts devouring her babies. Anyway, I concur. The nipples do take some tweaking to get right, but when dialed in this lora can really bring back some mammaries.
@Ponder_Stibbons Later we sang; "Thanks for the mammories..."
My results are very inconsistent, so I think I must not be using it correctly. Could you post an example workflow? I couldn't get one from your videos.
Shut up and take my money!
You gotta do a version 2!
What would you like to be added in v2?
@CaptainWebSpinner Is nipple growth possible?
@CaptainWebSpinner You have an awesome base. A couple other positions besides standing only and trying for more consistency would send your already awesome lora into a key spot!
Gotten some great results from this. Still trying to figure it out. Get some inconsistent results but I have been mixing it with other things. Am enjoying it though!
Could you make anime version of this?
It works with Anime Images too
@CaptainWebSpinner Could you try this one? https://civitai.com/images/76756620
Then send it via Google Drive, okay? I tried on site, it didn't inflate even bigger, it shows very weird stuff.
why lora is not available to generate in tensor
It was set as t2i model, thanks for informing, you can use it now
@CaptainWebSpinner what lora weight to use
@Xhavochero956 use lora weight: 1 and Cfg 6.5
in tensor is so hard to use your lora i can't work with comfy
Yes, They don't allow to adjust CFG, this lora needs 6.5 cfg 😕
you can try to make genderswap concept video wan
with Wan2.1-Fun-14B-InP_fp8 your lora is working on tensor
Yoo, you might want to retag your lora. They updated the tag and a lot of good loras, like yours, got omitted in the filter. Nice work!
Done, Thanks for informing 🙏
the only issue i have is that i want to use the last frame of a vid that i had added the lora to and it seems to not want to grow the breast larger than they were . how do i get hyper breast growth
This is well done! Works on both Anime and Real life.
how to stop them growing at some point? can I set a size? :D
You could maybe try combining it with one of the breast size loras? Run them on slightly lower lora strength? Haven't tried it but it might work. :)
I seem to frequently have generations with this lora where it shrinks the breasts instead of growing/expanding. Even if I only use the recommended prompt. When it works, it seems to be decent, but it's a bit of a pain how hit-or-miss it can be. My only thought is that it might be struggling with images where the breasts are already on the larger side. Hopefully this lora will eventually be updated to avoid this problem.
Yes you are right, it has problem when the breasts are already large, I'm trying to explore ways to make it work on large breasts too in future updates
100% can not ever get this Lora to do as it should. I've adjusted shift, steps, Lora strength, tried every prompt method I can think of, samplers, schedule, never works. Just makes a little jiggle of skin and that's it.
Are you using 480 or 720?
why your model is deleted from tensor
Tensor has hidden all nsfw models, they are doing some adjustments
If you are willing, would you please share some of the LoRA generation parameters you used for your motion LoRAs like this one? I have been trying to train a Wan LoRA with Civitai's trainer using only videos and I'm not having any luck. I have a decent dataset, but what I'm missing are numbers for the trainer.
Number of videos, number of epochs, number of repeats, how many in a batch, Network DIM Learning Rate, Network Alpha.
Thanks for any help you can give.
Civit trainer doesn't have option to select frame rate and frame sample parameters, So it's not possible to train motion lora without those parameters
CaptainWebSpinner Well... crap. Thanks for the help and the information.
CaptainWebSpinner On the chance I'm able to put together a Wan lora trainer that runs on my hardware, are you able to share some of your parameters? It would likely save me a lot of time with trial and error since lora generation at home is so slow. There's practically no information online about video/motion training for wan that I've been able to find. The online examples are all about still images. If you have the time and can share that data, I'd appreciate it.
darkroast175696 15 to 20 videos, Epoch and repeats any combination which results in 2500 to 3000 steps (number of videos × Number of Ecophs × Number of repeats = Total steps), Frame rate 16, Frame Sample (Number of slice) 5, Unet lr 0.0001, Network dim 64, network alpha 32,
CaptainWebSpinner That's great info to have. I haven't used a network dim/alpha combo like that with Wan yet. Thanks!
Works good on WAN 2.1 but not on WAN 2.2 It would be great if you could retrain it for 2.2 so we can use the higher resolution and quality.
im having trouble with this one.. the video goes very dark and doesn't work like it used to. is anyone else having this issue? it used to work very well
You mean on the civit genrator right? I checked i2v it's working fine for me
CaptainWebSpinner hmm, ill try again
Got the same problem on ComfyUI
I am also having issues. The videos I'm getting just slowly phase out the original image.
@FormidableFantasy found out that if I removed other LoRAs it started to work again.
@Deletetdyslexi it's the only lora I'm attempting to use unfortunately 😕
I'd love to see this on wan 2.2 <3
Tested it and it runs great with both the TittyDrop and Rip/Tear loras (if you can get your hands on them :/)
Aww it doesn't work too much with Wan 2.2, unless you crank it up around 1.5. More than 2 didn't make much of a difference... the result is better with Wan 2.1 anyway. It's fun, thanks ^^
What workflow do you recommend to run this Lora ?
Not really happy with this, doesn't seem to work
Details
Files
96774d10-5c2a-4904-862e-215d573300e5.safetensors
Mirrors
96774d10-5c2a-4904-862e-215d573300e5.safetensors
96774d10-5c2a-4904-862e-215d573300e5.safetensors
Breast-Expansion.safetensors
96774d10-5c2a-4904-862e-215d573300e5.safetensors
wan_Breast Expansion.safetensors
96774d10-5c2a-4904-862e-215d573300e5.safetensors
96774d10-5c2a-4904-862e-215d573300e5.safetensors
96774d10-5c2a-4904-862e-215d573300e5.safetensors
96774d10-5c2a-4904-862e-215d573300e5.safetensors
96774d10-5c2a-4904-862e-215d573300e5.safetensors