Version 2.0 Updates
Version 2 includes both High and Low Noise models, but the Low Noise model is optional. If you use both models, you will get a slightly different look, but most of the work is done in the High Noise model so that one can be used by itself if you prefer. The sample videos for the High Noise model were made using ONLY the High Noise model. The sample videos on the Low Noise page were made using both High and Low Noise LoRAs.
The training set included all the original video clips plus lots of new clips that were made using version 1, making a sample set of around 100 videos in a wide variety of scenes. This version did include some faces in many of the videos, so there will be more face bias in this version. You can work around that by only using the High Noise model and combining it with a Low Noise character model to change the face of the main character in your video. This version has also been tested with image-to-video and it works well there. You can safely use both the High and Low Noise files in I2V and it won't change the face of the person in the starting image.
The dataset included a few clips with changing hair, so that will work a little better in this version. I also tried a butt growth video just for fun and that seems to work, too. Overall, this new version does a better job of keeping the character's clothes on unless you specify something different. However, wan still has a habit of putting nipples in weird places once a woman's breasts get very large, so that's still an occasional issue.
Important: This model is 100% fictitious and is not based on images or videos of any real person. In fact, every video in my dataset was created in AI with the heads out of frame to avoid biasing the faces. No images or videos (neither actual nor AI-generated) of any real person were ever involved in the making of this model. LoRA dataset available on request.
Version 1.0 Notes
I know there are a couple other similar LoRAs out there (here and here), but I've been wanting to make a model like this ever since AI videos started to become a thing, only I didn't have the hardware to create full motion LoRAs. Now, thanks to a musubi tuner installation that I discovered from this model page, I was able to create my first Wan 2.2 motion LoRA. Finally!
Trigger word: brstgrw ("breastgrow" with the vowels removed)
Helpful phrases: "Her body is slowly changing", "Her bust slowly grows bigger", "Her breasts slowly grow bigger", etc.
Strength: 1.00 in High Noise ONLY
Pair with the Bursting Out High Noise model at about 0.75 if you're doing wan 2.2 image-to-video, and it will boost your chances of the woman growing inside her suddenly too-small clothes.
Where is the Low Noise model?
There is no Low Noise model.
At least not yet. Why? Because you don't need one. All my test videos using only the High Noise model are looking amazing so I saw no need to spend the extra hours it would take to train a Low Noise model that might not even help. Please trust me on this. All of my sample videos were made with only the High Noise LoRA set to 1.00. I used lightx2v for acceleration in both high and low noise, and I used Boob Physics in low noise. That's it, no other LoRAs needed. There is only one exception, a video I included to show that this LoRA can mix with other action LoRAs, and the settings are called out in that video if you look at the details.
This model is very reliable, as long as you remember to prompt for growth. More prompt guidance below. The model is not picky about faces, races, or body types, so please do use this model to make your favorite ladies more busty no matter what color skin or where they came from. The model also works well with Image-to-Video and I included one such clip in the Sample videos, featuring Sophie hanging a Christmas ornament.
Negative points: Many of the women in my videos have navel piercings and I swear it's not my fault. None of my training videos included piercings of any kind (or tattoos) so that's all Wan's fault. It's more likely to show up when you use lightx2v acceleration since the negative prompts are ignored with CFG 1.0. If you really want to avoid the piercings, your best solution is to do full-length Wan renders without acceleration so you can set CFG higher and put piercings in your negative prompt. Next point: It's hard to control the size of the expansion. Changing the strength of the LoRA doesn't seem to be very related to the final size. Instead, dropping the strength is more likely to just make the animation less smooth. Maybe a low noise model would help here, I'm not sure. If you really want to reduce the effect, you can work around this by changing how many steps you give to the high noise vs. low noise. My sample videos are all 4 steps / 4 steps. If you go with 3/5 or 3/6 or some other uneven split, you can reduce the impact of the high noise body shape.
Prompt guidance
The more times you use the word "breast" or "breasts", the more likely you are to have naked breasts popping out of clothes. If that's what you want in your scene, then this is how you achieve it. If you want clothed expansion it's more tricky since you have to avoid that word. I use "bust" and "chest" and it seems to help most of the time, although that's not 100% reliable. Still, it's much more likely to give you clothed expansion if you use those words instead of "breast". Beyond that, just trust in Wan's language interpreter. It's pretty good at natural language and almost any description you can come up with will work. Be creative! And share your videos, I really want to see them. Seriously, post those videos here so I can see them too.
One final word: This model does work somewhat in Wan 2.1 as well, which means you can use it in things like InfiniteTalk and Wan 2.1 Vace and so on. However, it's not perfect, and that's the main reason I'm considering training a Low Noise model as well just to make that work better.
I hope you all enjoy the new model!
Description
FAQ
Comments (57)
It's sad that nothing reliable like this was ever made for Wan 2.1
There was one, but it never works the way it's supposed to.
We moved too quickly from 2.1 just when the workflows were starting to get better, it got abandoned.
I know what you mean. I'm also working on a low noise model with this dataset since a low noise version will work in wan 2.1 videos as well. I'm new at wan 2.2 lora training, though, so it might be a bit of a wait.
@darkroast175696 ah nice, I didn't know that the low noise 2.2 bits worked as 2.1 LORAs lol that's pretty neat.
@jb8892 They're not always as good or accurate as the ones built from the 2.1 checkpoints, but they do work. I've used some different 2.2 low noise models in wan vace and infinite talk already and they appear to do just fine.
Oooo... clothed BE. I am very excited to use.
I am still waiting on one trained in ripping cloth :) :) :)
Thanks! And maybe I'll be able to do that someday, but I have to be honest it's gonna be really hard to train.
This is incredible, thank you and keep up the great work!
Glad you like it, and I'm looking forward to more of your content.
@bebobular I have some good news for you: https://civitai.com/images/114299526
@darkroast175696 thanks but the link appears broken - I get a 404 on it.
@bebobular Try again. I had just posted it and it looks like it's publicly visible now.
@darkroast175696 I can see it now - that looks very good. Looking forward to getting my hands on it! Have been trying to trick the BE Lora's into do it, nice work!
Wait - did you just do it with your Breast Expansion Lora and the prompt, nothing else?
@bebobular yes :)
Like I said in the video comment (I included the prompt) it isn't 100% consistent because some of my other tests didn't work as smoothly, but with some prompt practice you could probably make it more reliable.
Nice - I'm excited to give it spin
@bebobular I have another new(ish) lora that you might get some good use out of: https://civitai.com/models/2291854?modelVersionId=2579041
@darkroast175696 Awesome - it looks great, thanks for making a lora for Z image - I find it works stupid well!
@bebobular FYI, the latest release of Busty Women Low Noise model works very well in infiniteTalk. I've found that most wan 2.2 low noise loras will run in a wan 2.1 workflow and that appears to be true here. So if you want to take advantage of any of the motions provided in the new busty women lora (https://civitai.com/models/1670104?modelVersionId=2623123) then it should work for you. You can use it at strength 1.00 and it will replace boobphysics if you're using that in your workflows.
Yes, a model that leans toward clothed/stretching clothes, thank you! Can't wait to try it.
Yeah okay I used this in a render I've been struggling to get right and the first try was exactly what I'd been trying to do over 40+ renders, you are a legend.
@SinglePage That's awesome, I'm glad it's working for other people besides me. :) I'm looking forward to seeing your new posts. Oh - and thanks very much for the tips!
Very good. I've used the two mentioned in the description before, but they changed the character's face, making the lips thicker. This Lora changes very little.
Thanks for the compliment, and I hope you can get some good use out of this model.
Very interested in trying this out, completely new to comfyUI and i2v im trying to figure things out.
Before i try making my own workflows i wanted to learn from one thats already done but most of the vids posted even the examples if i download and drag them on comfy it says "unable to find workflow" im not sure what am i missing here or most of them just doesnt have the workflow.
I don't know why that happens with some of my videos. I know the data is embedded because I restore workflows from my old videos all the time so I don't know what's happening when I upload them to civitai.
Anyway, if you drop a link to a video you want to use as an example, I'll send you a link to a .json of the workflow so you can use that instead.
@darkroast175696 that would help a lot thanks! https://civitai.com/images/113850703 this one or any other of the example would be fine :), and yea ive been looking at workflow guides and other people also have same issue losing metadata when uploading/downloading vids, no idea why that happens as i just started using comfyUI and im completly lost :D but examples with json files all worked fine
@darkroast175696 I would also love a json, I'm still figuring out comfyui.
@robotfromspace I dropped you a link in chat.
What an OUTSTANDING Lora! So fun! THANK YOU!!
Looks super cool, but I can't seem to add it to image-to-video, it only allows me to find it in text-to-video on civitai.
True. That's an old problem with civitai's model selection system. They try to filter out all the thousands of models that wouldn't work in your video generation, like all the sdxl loras for example, but when wan came out it blew all that up. Wan loras are very flexible and you can use them in places where you wouldn't expect and civitai's system doesn't handle it well. Nothing much either of us can do about it, unfortunately.
@darkroast175696 Ah, shame. Sucks to hear but not your fault. I just don't have the setup to do local generations atm.
THANK YOU! Really love this lora and it works outstanding for me compared to some others I've tried!
Is there any chance to also get a reversed version, that can do shrinking? I know its a bit of a niche request
Thanks, I'm glad to hear the model is working for you! I have to say I'm not likely to do a reduction version since it's really not my thing. You might be able to make your own videos in this style by using this LoRA with simple backgrounds and then reversing the movie clip. As long as you don't have people in the scene walking or other motions that look obvious when played backwards, it might work for you. You'll just have to pay attention to the prompting for things like face expressions and so on.
@darkroast175696 no worries, that was going to be my next step but have never trained a lora before so thought it was worth a shot!
@darkroast175696 if you don't mind me asking, would it be possible to share your training data set?
@kemoihannoifri7767823 Sure, I can do that. I'll put a zip file together for you later and will send you the link in chat.
@kemoihannoifri7767823 I'm sorry I didn't get to this sooner, but I just dropped you a link in chat.
Works really well for I2V also! I found it works good in conjunction with https://civitai.com/models/2148922/large-breasts-strained-shirt if you want to keep the clothing on.
I've been messing with LTX-2 I2V model also, which has a lot of promise but very few loras so far. I'm going to try and train a few of my own but it would be great to run your training data against LTX-2 to see how it works. I found a training guide here https://youtu.be/po2SpJtPdLs?si=bTOUUM5j0DshxvXA
I've made some image-to-video clips with this lora, but I didn't think to try the other clothing lora with it, so that's a good suggestion and thanks for the tip.
I haven't ever tried ltx-2 so I'll have to take a look.
Yes, if you combine this with large-breasts-strained-shirt you get good results where the shirt/bikini stretches but does not come off. It's similar to what your burstingout lora does but it seems more predictable with normal clothes. I still haven't figured a way to get the shirt buttons to pop one by one, or the fabric to slowly develop rips and holes though.
As far as LTX-2 I have only been playing with it for a day using the template in comfyUI and am only getting one good result in ten. But it can generate a 5 second 720p video at 24fps in about 80 seconds with sound and lipsync, which is pretty fast! Plus it's open source and can run locally. It just needs more loras like yours and better workflows to be really fun.
ser can you share workflow? i cant get it to work with i2v
@clevenger You should also give my new version of Busty Women a try. See the comment I added to this Breast Expansion model and the videos I posted. They're all image-to-video using that new lora to help with the final look as well as guide the motion and so on.
@ponystalk69990 I haven't found a great workflow yet, I've been using the one in comfyui templates which works, but if you are going to generate large or long videos it really helps if you replace their VAE decoder node with the tiled VAE decoder. With long 720p videos it fit in 24gb of VRAM until the VAE decode step, then it jumped to 40. Doing a tiled VAE Decode dropped that down to under 24gb and really sped up the final part of the generation.
To anyone who's making Breast Expansion videos... It looks like my new version of Busty Women (https://civitai.com/models/1670104?modelVersionId=2623152) works incredibly well with this lora. You just put Busty Women at 1.00 in both high and low noise, set breast expansion to 1.00 in high noise, and describe what you want. The character won't bust out of her clothes unless you prompt for it. You can specify different sizes using the prompt guidance in the Busty Women lora, and you also get full use of the motions and actions. You don't need boob physics. I've just made about a dozen image-to-video clips with this setup and very simple prompting and it has worked every time.
Also, you don't have to be afraid to use the word "breast" - it won't cause wan to pop a boob out if you use the word. Give it a try!
Does this not work with GGUF Q8 wan files? I cannot for the life of me get the subject to remained clothed with any combination of the recommended loras and any type of prompt.
I'm pretty sure I've used it with gguf base models in the past. I agree sometimes I have to fight to keep the clothes covering the character's chest, but it usually isn't as bad as you describe. If you can share a workflow with me, I can give it a try on my end and see if I can make it work, maybe later tonight or over the weekend when I have some free time.
@darkroast175696 I'd appreciate if you gave some input when you have a chance! https://pastebin.com/WNqLtFYe
This is close to the workflow I'm using, the only differences should be I've added extra loras for lightning loras and your busty women / bursting out loras when using the stock Q8 gguf wan2.2. I prefer to try and use the nsfw fast motion camera adherance prompts on here because the motion is much better, but I'm worried they were causing the nipples to be shown.
I'll continue to experiment with the weights and combinations!
Also, was wondering what happened to the tit drop lora? Any plans for another version in the future?
Thank you for all your work!
@lumpy700 It should have been obvious to me at the time, but it seems the tit drop lora is also against the terms of service for civitai because you can use it to turn images of regular people into images of topless regular people. So that's why my search for other tit drop loras came up empty. So... no new versions, I'm afraid, it's a no go on civitai.
@lumpy700 I did a few test runs with a couple of different images of clothed women, one in a blouse and another in a t-shirt. I used a slightly different gguf because I had it installed and downloading new ones takes forever. The one I used was a Q6_K instead of your Q5_K_M, so it shouldn't be different enough to matter. I also disabled the unload model nodes in your workflow because I didn't have those installed and I'm assuming that wouldn't change the output at all.
Next, I tried a few different combinations of changes:
* different lightx2v loras
* different total frame count (your workflow says 49, I also tried 81)
* different step count (your workflow is 2 + 3 = 5 total steps, I also tried 4 + 4 = 8)
* different prompt text
The only changes that made any difference were step count and prompt. I added the breast expansion lora at strength 1.00 to the high noise and ran all those different options at 480x720.
My first prompt:
brstgrw. a woman wearing a t-shirt and jean shorts is posing for photos in front of a plain background. her breasts slowly grow bigger.
With only 5 steps, the expansion was not much at all. I bumped it to 8 steps (4 and 4) and it was a much more noticeable change, and she stayed clothed during the expansion. Then I tried to break it with an extreme prompt like this:
brstgrw. a woman wearing a t-shirt and jean shorts is posing for photos in front of a plain background. her breasts slowly grow bigger. her breasts grow very huge and round, stretching out her shirt. at the end of the scene her breasts are gigantic. she grabs her breasts with her hands and smiles happily.
That made it break badly. Her breasts popped out of the t-shirt and the video looked terrible because this model does not do I2V naked breasts well at all. I tried to find a middle ground and this prompt gave me a stronger reaction (bigger bust) but kept her clothes on:
brstgrw. a woman wearing a t-shirt and jean shorts is posing for photos in front of a plain background. her breasts slowly grow much bigger. she grabs her huge breasts with her hands and smiles happily.
So I guess my suggestions are these if you want the character to stay clothed and covered:
* pick a good starting image. If she's wearing a strapless dress and her chest is barely covered in the first place, you have less chance she will stay covered in the video.
* use at least 8 total steps with lightx2v. I've gotten good results from 8 and 10 steps, and lately I've been using the Cache DIT node for Wan to speed up generation and it's like magic. It lets you skip some of the steps, which basically lets you boost the number of steps and still save time. I really recommend it. See my published Wan Daily Driver workflow for notes on that node including where to get it and how to install.
* prompting. It looks like if you go overboard trying to achieve cup sizes in the back half of the alphabet, you're probably gonna bust out of the clothes. It also sometimes helps to avoid using the word "breast", so I often use "bust" instead. Try starting with "her bust slowly grows bigger" and you can also try "her bust slowly grows bigger inside her clothes" or "under her t-shirt" or similar wording.
That's about all I can suggest for now based on the workflow you sent. Good luck!
@darkroast175696 Thank you so much for taking the time to run the tests! I did increase the step count to 10, and moved away from the NSFW gguf files because they are much too biased towards nudity. My results are much more consistent now! thank you!
Also sorry one more question, what frame rate and interpolation multiplier do you target for your videos? Mine are always kind of slow.
@lumpy700 I actually kind of like the slightly slow videos, but it's up to you. I generally set my videos at 3x interpolation and 32fps.
@lumpy700 great! I didn't even think to ask if it was a custom gguf. I'm not surprised the nsfw models gave you the results you were getting. Glad you solved it!
hi, can you publish that model as i2v model? i want to choose it when upload the video with another i2v loras. very good lora btw
Thanks, I'm glad you like the lora! It might be better to make a real image-to-video version of this lora. I've been intending to do that with the busty women low noise lora too, in order to fix some of the weird anatomy you can sometimes get in image-to-video workflows. I'm not sure when I'll get to it, but I'll try and work on the breast expansion model soon.
@darkroast175696 that sounds good
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.