The purpose of this LoRA is to generate more interesting breast play trained off the Wan2.2 base model. Please be kind in the comments. Constructive criticism is always welcome as it helps others including myself. Also, positive feedback is much appreciated <3
Version 2
Features
Reduced Face changing issues compared to Version 1
Additional breast play with another subject involved. Original dataset included, too.
Trained on close-up and full body
Trained on different style of breast play: grabbing, lifting, covering, pinching nipples, and holding
Trained on breast play with single breast play: right breast and left breast
Tips
Tested on High Str 1 and Low Str 1 to 0.8. Appears to work with 1 strength for both.
Increase High Str > 1 to enhance the breast play and overcome issues described below. I tested at 1 and 1.5 for High, works well. Leave Low between 1 and 0.8
Prompts
she grabs both her breastsshe continues to grab both her breaststhe man is grabbing her breastshe is covering her breastsshe rubs her breastsshe grasps her breastsshe grabs her breasts with both hands and pinches her nipples with her fingers Lucky if you can get this ^^^ one to work, its awesome!
Make sure you are handling the high / low transitions correctly as it relates to the steps. Here is a k sampler that does it well: https://github.com/stduhpf/ComfyUI-WanMoeKSampler
Issues
Speed up LoRAs may slow down the motion. This is a KNOWN issue with the speed LoRAs. More information: https://huggingface.co/lightx2v/Wan2.2-Lightning/discussions/20 and Developer comments here: https://huggingface.co/lightx2v/Wan2.2-Lightning/discussions/26
Version 2's breast play is more subtle then Version 1. I reduced the training Epochs for Version 2 to resolve the face change issue but this seems to have made the breast play more subtle compared to Version 1. Try using V1 high and V2 Low or increasing the V2 High > 1 and keeping V1 at 1.
Shoutout to all the LoRA creators out there doing the insanely tedious work of tagging, editing and running training sessions
Version 1
(issue): Some users experience face changes with this LoRA. This is because the LoRA is overcooked. The LoRA is still usable. However, it means that the strength and workflow may need to be adjusted - review tips.
Features
Compatible with both T2V & I2V Wan2.2 Models
Trained on close-up and full body
Trained on different style of breast play: grabbing, lifting, covering, pinching nipples, and holding
Trained on breast play with single breast play: right breast and left breast
Tips
Try between 0.5 to 1 strength. All testing was at 1 strength on I2V with two LoRAs (speed + Breast Play).
Reduce strength on the LoRA if the face changes.
Consider using https://github.com/stduhpf/ComfyUI-WanMoeKSampler or adjusting High / Low steps 'correctly' https://www.reddit.com/r/StableDiffusion/comments/1mkv9c6/wan22_schedulers_steps_shift_and_noise/?tl=fr
My Wan2.2 Cover Videos
Workflow is T2I SDXL/Flux/Pony -> Image Upscaler -> Wan2.2 I2V -> Upscaler -> Add FPS
https://civarchive.com/models/1916073 (text2img & img2vid). Credit DonutsDelivery & Daxamur
Training
Trained on Wan2.2 T2V base model. 100 high quality images and 50 videos that were handpicked and farmed raised to perfection using musubi-tuner. This LoRA works with both I2V and T2V.
Adjusted training variables so that it works better with other LoRAs. Training was fined-tuned over 10+ runs for each model to get the optimal loss accuracy. I used different tagging descriptions for the dataset to improve variability.
Description
(issue): Some users experience face changes with this LoRA. This is because the LoRA is overcooked. The LoRA is still usable. However, it means that the strength and workflow may need to be adjusted - review tips.
FAQ
Comments (41)
Looks great!
But, please include metadata/workflow on your examples! Be like the great playtime_ai before he died!
@xabone3077154fdfdfd My workflow is a combination of several different workflows that I've run at different times (over the past 12-months). There isn't just "one" workflow. The metadata isn't saved when I move from the generated image workflow to the I2V wan2.2 workflow and when I complete the upscale/add fps process - sorry 😞. The posted videos are text2img (SDXL or Pony or Flux) -> Upscale/FaceDetailer -> Wan2.2 img2video -> Upscale -> Add FPS
SDXL images are from https://civitai.com/models/664292/donutsdelivery-sdxl-workflow-simplified-upscale-facedetailer-ipadapter
Some of my images are from this workflow https://civitai.com/articles/14582/comfyui-workflow-sdxl-facedetailer-upscaler-and-batch-selection
Some of the videos (Wan2.2 img2video) are from https://civitai.com/models/1895190/wan22-i2v-comfyui-workflow?modelVersionId=2145291
Upscale and add FPS https://pastebin.com/raw/8RcEvFGL
I'm going to try to organize all the workflows and post an article on the process. I'm trying to decide if I should create one massive workflow or break them up in parts like how I normally run them. If I use one workflow, then you can see all the metadata but it will be a MASSIVE mess. If I break it up, its not possible to include all the metadata in the posts but its really easy to understand and run. I've just been super busy doing these LoRA trainings - consumes a lot of time preparing datasets and testing.
@xabone3077154fdfdfd Here are the text2img and img2video workflows: https://civitai.com/models/1916073
@Kierkegaard420 I just had a look at your upscaling workflow, it's strange how you're using a still image to match the colors for a video. I don't think that should work but I'm gonna try anyway, for science. For matching image colors my favorite node is Image Color Match from comfyui-easy-use package, with the default wavelet method. It's really the best I've found so far but it couldn't handle a batch of images from a video, maybe there is a way
@Tetsuoo Yeah, to be honest with you, about 90% of the time I don't know what I'm doing...
@Kierkegaard420 It didn't work, as expected. Back to color correct manually, sigh. lol
This doesn't work well for I2V, it immediately alters facial likeness.
@psspsspsspssspss Try decreasing the strength. Also, make sure you are doing the high and low noise model steps correctly. All my posted videos are I2V without major facial issues so I'm not sure why you are experiencing changes.
For example, base image https://civitai.com/images/97526588 with I2V https://civitai.com/images/97448768
Let me know if I'm missing something or miss understanding.
@Kierkegaard420 idk if that second one was supposed to have a workflow or you were just proving the point. But in just a couple I2V tests I had almost perfect likeness with both at full strength. However, in general, when dealing with movement and physics it's High, and when dealing with style and quality and it's Low (for example, a nipple tuner would go on low, while this BreastPlay-high lora could be the only lora in the High slot and it would still work perfectly). You could however turn down the low lora to "somewhere" between 0.3 and 0.8. Sounds like someone liked 0.6. Since it's the low half, it should hopefully keep its "style and quality" opinions about the face at a lower strength, while retaining the movement and physics quality from the high lora.
This is just my experience. I don't know if this is documented somewhere for real. Maybe.
@makiaeveli Okay, thank you. I'm not really sure why some people are having issues and others are not having issues.
@Kierkegaard420 I've got no issues, as you can see from the ~100 videos I've posted so far. Half strength seems to work well (~.5)
@solss_ Thank you. I wonder if my LoRA is cooked to much. Are you using other LoRAs or just this LoRA by itself?
@Kierkegaard420 Just using half strength (.5) on both high and low for your two models, and lightx2v 480 Wan 2.1 lora at 2 strength on both high and low models since it doesn't introduce any slow motion compared to the 2.2 lightning loras. I'm using context options on Kijai's WanVideoWrapper and multiple prompts for different segments all in one video generation. That's what those | symbols are for. 81 frame segments at a total of 329 frames for twenty seconds. I would say 75% of my generations turned out great. No facial distortion or quality degradation.
@solss_ Thank you for the context.
Are all outputs slow motions?
It simply looks like most of the training set contains slow movement. I don't think the training data is at a different FPS than any other lora, although I can't know that for sure. In my tests the motion was slow, but it wasn't "slowed", if you know what I mean. Like look at the blinks and the muscle twitches, they're all pretty much normal.
@Visento My examples are slow motion because I'm using the Speed LoRA which naturally causes motion to 'slow down'. The dataset was trained on breast play and hand movement that was sensual, slow and smooth. You should be able to offset this by prompting correctly, 'quickly grabs breasts' etc.
@Kierkegaard420 @makiaeveli Got it, thanks.
Thank you! This fixed the weird breast physics I was seeing. I use it on strength 0.6 and do not see any altering to faces. Great job!
Thanks for the strength, I was trying all around that
Very good! No jelly titties, well done and thank you!
Thanks for the tips, that special MoE KSampler is making noticeable improvement in the results !
Wow, I used this as a breast helper, and it turned out great!
could you do a similar concept but breasts grabbed/squeezed and groped by another or by the viewer pov.
@obinna7713 Yeah, probably. I just need to build out the dataset for something like that. I will likely do a version 2 of this LoRA so I can add that to the feature list.
You can just prompt that?!
@Lora_Addict true
https://civitai.com/models/705447/fondled
Tried this recently in 2.2 (add it to both high and low, and mess with strength), and it works well, as an option. (The examples are synthetic, but I tried it with realistic scenarios and it worked great.)
None of that is to discourage a 2.2 solution, just letting you know you have an option now.
One of the best loras round here
changes face T_T
@brown_reason_to_live fixed. thank you.
Thank you for the LoRA and for considering file size in the equation.
both called 1.0
@kllrpenguin Someone pointed this out. I updated the versions a couple minutes ago but it seems like the site has cached the old update. Its the first 2 releases listed near the title. Sorry.
Please show some love to T2V, as well.
@honryindian This technically works with T2V because its trained on the T2V base model. I have not extensively tested it but give it a try. Obviously, its an issue with Civ to run but may work locally. I'll get working on a T2V release.
@Kierkegaard420 Why is it registered as I2V base model then?
@wewewew Gotta set it like that for people to be able to use it with the on-site generator on I2V. In my testing it performs much better on I2V than T2V regardless so I2V is the way to go.
Well, tried this lora, with "she grabs her breasts with both hands and pinches her nipples with her fingers " prompt, nice results both clothed and without.
And then I noticed that not turned this lora "on". -___- Anyways, with lora works better.
V2 is more subtle then V1. It definitely works as shown by the LoRA comparison cover post which shows the LoRA on and off....
Details
Files
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Mirrors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
BreastPlay_HIGH.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BP-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BP-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2-I2V-BreastPlay-HIGH.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
LORA_I2V_Breast_Play_2_H.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors
Wan2.2_BreastPlay-v1-HighNoise-I2V_T2V.safetensors