Wan 2.2-5B:
Trigger words: Br34st 3xpan510n
This new version of my breast expansion lora has been trained for Wan Video 2.2-5B. Why? I don't know, I suppose it was just the first model I tried to train a lora for after a long time away. It seems to have gone fairly well? This model is much (much) more consistent at making breasts grow than the Hunyuan version, but I've really only tested it with image to video. It does fine with animated content or photorealistic stuff.
As far as I can tell, there are two main drawbacks: It does terribly when it has to generate nipples from nowhere (either because they were covered in the original image or because they get covered for more than an instant during the course of the video), and it doesn't do well with loose clothing. If you try to grow some boobs in an oversized hoodie for example, the hoodie itself will sprout breasts. Are these things that a v2 might improve upon? Honestly, yes. I didn't actually expect this training (which I did locally on a 4070) to work at all, so I didn't really put to much thought into curating the dataset. That being said, don't expect a new version anytime soon (or even ever, perhaps).
Having learned my lesson while posting the Hunyuan version, all of the videos posted in the carousel aren't upscaled, inpainted, or otherwise altered from their initial generation in any way, and all were generated locally using the aforementioned 12GB 4070 (and quite quickjly too, as each video taked only about 3-4 minutes). If you need to start with respect to a workflow, I co-opted this workflow by JCB2.
Anyway, thanks for giving it a try if you do. I really enjoyed seeing all of the stuff that you got up to with the Hunyuan version, so here's hoping that you're inspired to make some more cool stuff.
Hunyan Video: Hello everyone! Sort of a good news/bad news situation here. Good news: boobs will sometimes get bigger. Bad news: it's pretty inconsistent, I can't get it to play nice with other LoRAs, and I have no idea what I'm doing.
This was trained on a dozen or so small BE clips (256x256, 33 frames) using Musubi Tuner. I've never trained a LoRA before, and I really wasn't expecting to share this one, but I wound up getting a few (a FEW) good generations out of it so I thought I'd share. If anyone decides to give it a go, I'd recommend the following:
Keep the strength at around a 0.8. I have gotten results anywhere from 0.6 to 1.0, but any higher than that and things will go south real quick.
Use the keyword Br34st 3xpan5ion at the start of your prompt. Other than the keyword use natural language. Explain what it is you'd like the breasts to do as they get bigger. Burst out of clothing? Push a bra down? Bounce and wobble? I don't know if it actually does anything, but it makes me feel better if the gen turns out all right and I prompted it that way.
And... I dunno, have fun I guess? I look forward to seeing what all of you talented people can make of this.
Description
FAQ
Comments (35)
Great Lora. How many epochs did you choose? And about the clips? Have you trained in 12GB VRAM?
Thanks. This is the 70 epoch version. I ran the training for 100 with a save every 5. This one just seemed to give the best results. And this Lora was trained on a 12GB 4070. It took a bit more than 24 hours.
Not sure what you're asking about the dataset. I chose a bunch of short gifs with relatively smooth growth from my private stash. I captioned the first frame with Joy Caption, and did a very light edit where needed.
@jkbutcher1015940 Many thanks for the answers.
@jkbutcher1015940 Thank you for sharing this and OMG thank you for this lora. Amazing!
LOL Busty indeed.
Holy boobas Batman! Awesome work. 😉
now this is epic
This is a fun lora to play with, thanks!
Can you post the workflow you are using? My videos are coming out a bit blurry and would like to see if you are doing something different than me.
Thanks for playing around with it. I've added a new video to the carousel that has the workflow embedded in it. Just right click the Asian woman with the bathrobe and save the video to your computer, then load it into ComfyUI. If you save any of the other one's you'll just get the quick and dirty upscale workflow that I used on them before posting since I disabled the upscale and interpolation in the main workflow to iterate faster. I hope this helps, but I've a feeling that videos produced with this Lora are never going to be real lookers given the low resolution training data. I tried adding in a few high resolution still pictures along with the small vids, but that was several attempts ago and apparently was when Musubi Tuner had a bug that caused severe tiling artifacts (or something... have I mentioned that I really don't know jack about any of this?), and so I just tossed all of those Loras after just a few gens. Maybe if I do a v2 I'll try again now that I have a better baseline.
Growth of the breasts is almost directly proportional to the length of video created. (Longer video = bigger breasts) Which makes sense because in the training videos the breasts got bigger in each frame.
When I get a 5090 I'm gonna have the thing running overnight making these.
sorry i just drag and dropped the video on comfyui but didn't gave me anything, could you share your workflow in other ways pls?
Sorry that it isn't working for you. The workflow is just this one from Bonetrousers: https://civitai.com/models/1092466?modelVersionId=1243590
The only thing you'd need to add would be an Hunyuan Video LoRA Loader node in between the 'Load DIffusion Model' node and the 'Set_Model' node.
You gotta do a V2!
I certainly won't rule it out, and in fact I've been pulling together some stuff for a new dataset. But I'm not sure if it would be any better in the end. There's two main limiting factors: I am at the very end of my technical know-how, and I only have a 4070.
I think the best way to improve would be to use higher resolution videos for training, but I simply can't do that. And while it's possible that I could use a cloud service, I am pretty deadset against spending (or making) any actual money. Were I to start doing that, I think it would change the nature of this little hobby of mine irrevocably.
@jkbutcher1015940 How long would a training run take you on say an A100? You get roughly 10h for free on Modal (per month). You just need a little Python to use it. If I remember correctly, "pip install modal", followed by "modal setup", and then you should be able to run your Modal Python scripts (.py) with "modal run <file>". This is what I used for my Illustrious LoRA, trained with OneTrainer (config file and dataset folder should be in the same directory as the script):
After training is complete, I download the model with:
modal volume get Output "OneTrainer/models/lora.safetensors"
Hope this helps.
Please add better trigger words to create decent results with only the trigger words
When I load the workflow the only thing that does not appear is EmptyHunyuanLatentVideo. Where do I find this?
Solved the issue by updating comfyUI
Here is how I got this to work for me. I updated comfyui and pytorch in the comfui update directory. I then downloaded the video of the Asian woman as described by OP. I opened it in comfyui to view the workflow. I took note of the diffusion model, clip loader, and VAE used then googled and downloaded them from civitai. Next I plugged them in and left the existing settings the same but lowered the temporal size to 40 from 64 on the VAE decode because I only have 8gb of vram and I was getting memory use errors. Suddenly I was getting good quality videos. Hope this helps someone.
Update: Its taken me lot of tries to be able to get this to produce a couple of positive results as OP stated.
Okay, here's a quick Tutorial for everyone who is having Problems:
Download the Video (at the top) with the Asian Woman in the bathrobe. Drag and drop that video into ComfyUI and it will open up the Workflow. You might have to download some missing custom nodes, you can fix this by using the ComfyUI Manager and going to "Install Missing Custom Nodes"
After that, you should download these:
- The Breast Expansion Lora itself (ComfyUI\models\loras)
- Clip_l.safetenors (ComfyUI\models\text_encoders)
- llava_llama3_fp8_scaled.safetenors (ComfyUI\models\text_encoders)
After that you need to select them in your workflow and you are ready to go!
actually works fairly well once you understand the basics. so i redact my original comment ;)
Please add the wan version, thank you
I'd love to be able to give the people what they want (especially since I'd really like it as well). I have tried many times, and while I do have a nominally working Wan 1.3 t2v LoRA, everything it gens looks terrible. And unfortunately, I am unable to train on any of the larger models given the hardware I'm working with.
@jkbutcher1015940 Thank you for your work and replying to my message. I am sorry for the difficulties you encountered in making the wan version. I hope that one day in the future, you will successfully make the wan version. I am looking forward to it.
I believe i'm blind, i can't seem to find the Asian woman in the bathrobe?
The seventh clip in the carousel I believe, right after the one with the magic wand. I suppose on second glance it isn't actually a bathrobe, just a terry cloth cardigan or some such.
@jkbutcher1015940 Apologies if i reply twice, can't seem to find my first response. It says unable to find workflow once i drag it into comfyui, tried converting video to webp and still no joy, it's worked on others so bit confused why its not working on this one.
@artgen44 I just want to make sure we're on the same page here. You're opening the video here on Civitai, right clicking on it and saving the video to your computer, then dragging the file from your downloads folder into ComfyUI? I don't see why it wouldn't work. It should be an .MP4 file when you save it, and should remain as such. All that being said, there's nothing particularly special about that workflow, and any Hunyuan workflow on this site that allows you to set a LoRA should work fine.
@jkbutcher1015940 I believe when you upload a mp4 it scrubs out the actual workflow from the file. I have been trying to do this for a while and I believe you need to export it as a workflow.
Could we have this for Wan also
I'm just starting to study the image to video, tell me how you use it? Hunyuan Video requires a lot of VRAM
please upload your workflow as a separate file. It is not hard to do.There is no bathrobe women you mention in the description and I have tried every video with no success of importing from the meta data that you claim is in the video.
Same here, please add the workflow :)
this is glorious! I love it! could i beg you to make a tiddy drop one?