I've done plenty of test runs but CIVITAI takes extra long to verify my posts now. Please check back soon!
May 23 2025 Update — I’m Back 🎉
Hi everyone! After a short hiatus and getting familiar with Civitai’s new guidelines, I’m finally back—and I’m dropping two fresh checkpoints for the Twerk LoRa.
V2.0 Free through Civitai
Exclusive Access: EP100 for $1.5 https://singularunity.checkoutpage.com/sutwrkep100
V2.0: Subtle but mostly Stable
What’s new? V2.0 tones things down from the punchy V1.0. It keeps the natural jiggle and muscle dynamics you loved, but with cleaner geometry and far fewer surprises in general.
Why you’ll like it: If you value predictability and more subtle results.
Exclusive Access: Epoch 100 — “Wild & Dynamic”
What’s different? Think of Epoch 100 as the adrenaline-fueled cousin. It can push motion intensity and deformations further than ever—but that can come with extra instability.
Pro tip: Bring patience (and maybe a few extra GPU cycles). With the right prompt-tuning and seeds, you can squeeze out some truly jaw-dropping animations.
A Quick Word on the New Rules
Civilta's policy changes took me by surprise at first with nearly 90% of my NSFW posts were suddenly hidden, but after digging in, I see where they’re headed. And tbh I think in the long run it will be a net positive.
April 2, 2025 Update
Hello everyone! I’m excited to introduce my Twerking LoRa model, specifically refined to capture natural, physics-driven twerking movements. Building upon previous feedback and learnings from previous models such as MotionCraft, this version focuses on the subtle jiggling, bouncing, and muscle deformation that bring twerk performance to life. As always, I rely on your results and feedback to keep improving—so please share your outputs!
Note: This model works great with the MotionCraft Lora!
Image Selection Guide
Realistic Female References
The LoRa has been trained primarily on realistic human figures. For best results, use images that clearly display hips and thighs.Good Lighting
Even lighting highlights soft tissue and muscle engagement. Dim or uneven lighting can reduce visible motion details.Minimal Background Distractions
A simple or uncluttered backdrop helps the model focus on the twerk movements.Body-Hugging Outfits
Nude, Shorts, leggings, or tight skirts are excellent choices—loose but form-fitting clothing reveals the bounce and elasticity of the thighs and glutes.
Trigger Prompts
Incorporate the word twerk or twerking alongside other descriptive phrases to elicit the best motion responses. Here are some examples:
“Twerking motion with natural jiggling and bouncing”
“Dynamic twerk emphasizing elastic thigh and glute deformation”
“Rhythmic twerking emphasizing subtle physics-driven movements”
“Steady twerk routine with visible muscle contractions and releases”
“Twerking performance showcasing realistic skin stretch and recoil”
Example Prompts
Below are five example prompts resembling the style you requested. They all include the key term twerk or twerking, with richly detailed descriptions emphasizing the body’s rhythmic, elastic movements.
1. A woman wearing a bright green sports bra and matching biker shorts stands upright, hands on her hips. As she begins twerking, her glutes contract in a fluid, rhythmic motion, producing gentle jiggling in her thighs and lower back. Each bounce highlights the elastic recoil of her muscles and skin.2. A woman is twerking, moving her hips and thighs in a rhythmic manner. A smooth, continuous, and cinematic camera pan shows her body moving in a fluid and smooth motion, with her muscles contracting and relaxing. The subtle physics-driven soft tissue dynamics are visible as her skin stretches and rebounds in response to her movements.3. A woman performs a captivating twerk routine, showcasing her natural jiggling, bouncing, and the elastic deformation of her thighs, glutes, hips, and lower back. With her blonde hair styled in braids and a striking tattoo on her upper arm, she bends over and lifts her legs to reveal her butt in gray tights. The scene captures the fluidity and grace of her movements, set against the backdrop of a cozy bedroom with a bed, a window with blinds, and a desk with a chair.4.A woman in a white ribbed crop top and olive green shorts stands with her hands on her hips, showcasing her toned back and fit physique. As she begins to twerk, her natural jiggling, bouncing, and the elastic deformation of her thighs, glutes, hips, and lower back regions are highlighted, capturing the subtle physics-driven movements that define her dynamic performance.Description
WIP LoRa
FAQ
Comments (33)
I2V :'(
Can we please get a T2V version :)<3
Hi,
I'm planning to train a character LoRA for Wan2.1 (specifically the 14B model) using the Diffusion-Pipe framework on WSL. My goal is to generate short video clips featuring a specific person.
I have both still images and short video clips (e.g., 3 seconds @ 16 FPS) of the person. I was wondering if Diffusion-Pipe supports training a LoRA using a mixed dataset containing both image (type = "image") and video (type = "video") subsets within a single training run, referencing the same concept folders?
For example, could I configure my dataset.toml like this (simplified):
[[datasets]]
type = "image"
image_dir = "/path/to/my_data/"
subset_dir = "20_my_person/"
# ... other image params
[[datasets]]
type = "video"
video_dir = "/path/to/my_data/"
subset_dir = "20_my_person/"
# ... other video params
Use code with caution.
Toml
Or is it necessary to train exclusively on images or exclusively on videos when using a T2V base model for this kind of LoRA? Does Diffusion-Pipe handle merging or alternating between these different data types smoothly during training?
Any insights or examples of how to configure the .toml for such a mixed dataset would be greatly appreciated!
Thanks!
create separate directories for both. One images and one videos
does this work in tensor ai? How would you add it?
Excellent work! Thanks for this model. This is one I've been searching for. It works very well. I have a couple of questions:
How did you get the glutes to appear so supple and soft, and wobble so loosely? When I run it, the glutes jiggle but not as much. The glutes physics are tighter and more firm than the ones in your videos. What are your settings?
Why does the model keep trying to alter her position? I selected an image of a woman in a doggystyle pose. I want her to twerk in that position, but she keeps trying to stand up. How do I fix this?
Hey Solo, thanks for the generous tip. I've noticed the positional changing and it's probably due to the dataset I trained it on a combination of both standing and kneeling twerks but I did not differentiate enough between the two as I thought much of the context would be derived from the initial frame or image used. You'll have to run a few seeds to get the output you want, you could also try adjusting the prompts as well. When testing I stick with a single general prompt, so all the videos I posted used. "A woman is twerking, moving her hips and thighs in a rhythmic manner. A smooth, continuous, and cinematic camera pan shows her body moving in a fluid and smooth motion, with her muscles contracting and relaxing. The subtle physics-driven soft tissue dynamics are visible as her skin stretches and rebounds in response to her movements."
Hopefully future iterations I can correct this.
@SingularUnity Got it. Thanks, SingularU. I'll try that prompt, hopefully my physics will turn out the same. I wish I could get the exact workflow but it appears you use a Hunyuan for upscaling. My system doesn't like Hunyuan at all. I'm Using 12Gb VRam GTX 3060. The 12Gb should be enough for low VRam settings but it's not for some reason. I have to stick solely to Wan 2.1. You're welcome for the tip. I appreciate your work.
@SingularUnity SUCCESS! Your prompt works perfectly. 😁 Thanks. I wish prompting wasn't so complicated 🫤🤕. I mean, using my own prompts never turns out right. That sucks elephant ass.
Hey! I'm currently testing this out as well, and here's what I've figured out so far:
When you describe the pose clearly, like saying "A woman is kneeling on all fours on her bed for the entire clip while she is twerking". It helps prevent her from standing up mid-animation.
Appreciate it
Thank you
I think I m figuring this out a little. If I make a workflow from scratch and attempt to recreate a scene like this using a different LoRA, simply using prompts will not help, right? I would have to train the LoRA, correct? If that is correct, can this be done using only ConfyUI or must I use a 3rd party application? Are any of them free?
To train LoRa's I use diffusion-pipe it's free
So the answer is YES? Thanks for the resource.
What negative prompts are you using to fine tune the results you want to get
Just the standard found in my workflow
@SingularUnity I’ll be honest all I see is Chinese alphabet and the word JPEG in English. In the lower box of the WanVideo TextEncode Box
weirdly going on here, but yeah.
what?
SkyReels-V2 - Another Long-Length Video AI Model But With Wan 2.1 ! (Up to 30 seconds)
Nice one ! Thanks !
I'm using this Lora mostly with other Loras, but I noticed that when the strength is too high it produces very weird motion, so in general I use this Lora with strength 0.2 or 0.4. This is a good Lora! I can't train Loras but I'd be glad to contribute to a future update of this Lora
I'm glad you enjoy it, that's interesting to find. I'll run some gens and see ways I can improve the LoRa thank you.
Tried image to vid with a clothed woman. Got a "Adult Content" warning and blank image. Then a blank image.
Me too literally SFW and it said adult content....
@ArtByD YEah, I figured it out. Just CIVIT AI screwing us again. Part of their new rules. I had to go find it after wasting over 1700 buzz "NON REFUNDABLE" I canceled my subscription here and I went back to mage space for now. But I'm looking for something more consistent.
@Guam69 I like tensor art and the ability for controlnet
@ArtByD I really liked CIVIT AI until they started getting more and more complicated with the rules and enforcing them randomly. Then they started being random with what was available. I love albeddo base for when I am serious and not just making random oddity porn. But, I don't ever know if it is available from day to day. And I'm sick of "Image generation is down" messages. They have turned buzz into a bad game. Spending buzz to vote on what lora will be available? Screw them.
@Guam69 @ArtByD The payment processors "Visa/Mastercard" turned off their accounts and forced them to put more Moderation. That is why you can only pay with crypto now. They do this to podcasters and media they want to suppress as well, it's a problem.
@Bobby_Reed They started this a long time ago. And that does not change the fact that the image was within their rules. Which also has been going on for a long time. I mean, I had images pulled because the image kinda looked like a celebrity a little. But at the same time, some loras want to put a small child in every other generation even when you fill the negative prompt with every word you can think of to avoid that creepy shit. The credit card companies also, did not dictate the "NO REFUNDS" policy did they? I don't think mastercard said "Oh, and if they follow the rules, let AI still arbitrarily block the image and NO REFUNDS!! Muah hah hah"
@ArtByDThanks for the info. I have been combining Civit, Tensor and Mage space to generate some great video. I think I will like Tensor better once I am used to it. Mage Space went to crap when they decided to change the interface and I am still not happy with it. They made it both simpler and much more difficult at the same time.
Sorry folks, Nothing I can do about it, technically twerking is "Mature" themed so I'm hesitant to change that to see if it works and find out I'm breaching some element of the new rules. In the future I will release content that is more suited to Civitai outward taste. In the meantime this model works with both clothed and nude females. Hope you all enjoy
Just tried this with a woman wearing Lingerie and
got a Blocked Image Mature content restriction...
Details
Files
SU_Twrk_EP62.safetensors
Mirrors
SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
140_SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
WAN_SU_Twrk_EP62.safetensors
SU_Twrk_EP62.safetensors
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.