Introducing Mio.
Mio is a woman of Japanese descent in the grande tradition of Hitomi Tanaka and other busty Japanese models. She is my first Wan 2.2 character LoRA so it's not perfect, but I'm learning. See below for model and prompt guidance. The keyword for the model is simply "Mio".
Important: This model is 100% fictitious and is not based on images of any real person. She originated with images generated by AI and has evolved from there using only AI rendering. No images (neither actual nor AI-generated) of any real person was ever involved in the making of this model.
Mio used to be a geeky teenager who was into Pokémon, rebuilding her gaming rig, board games, and comics. Then she grew up, and out, and nothing else changed. She still loves all those things, only she's a lot more jiggly, her shirts are all stretched out, and she attracts a lot more attention than before.
Usage and Tips:
As a standard Wan 2.2 model, this LoRA comes as a pair with High and Low noise models. You should get good results with a standard 50% / 50% split between high and low steps, but feel free to adjust as you like. Mio sometimes appears with odd nipples with an extra bump or two. If you see this and you like the clip otherwise, try reducing the low noise model to 0.50, that sometimes helps. If you reduce the strength of low noise, or even turn it off entirely, you can reinforce the Mio appearance with a phrase like "a pretty Japanese woman with short black hair and and round wire-rimmed glasses". Even at normal 1.00 strength, she sometimes appears with longer hair unless you reinforce with the "short black hair" description. In other words, those extra descriptive phrases aren't necessarily required, but they sometimes help so it's easiest to just leave them in when working with this model.
All the samples were generated with a lightx2v accelerator, with either 4/4 steps or 5/5. The only models in use are Mio and Boob Physics, both at 1.00 in high and low unless noted otherwise. All the visuals and motions have been trained into the model.
In my experiments, the high noise model can stand very well on its own. With descriptive text about Mio, you often don't even need the low noise model. The face details are not quite the same, but you can get even more pure Wan cinematic detail if you don't use the low noise model. So give that a try and you might be pleasantly surprised by the results. I also found that the high noise model can work very well as a general-purpose "woman with huge breasts" model if you give other descriptions instead. You can give her long blonde hair, for example, or change other things about her and still get the general body shape to fit into your scenes as you like.
I hope you all enjoy working with Mio and please do post your video creations here in this model page so I can see what kinds of adventures you have in mind for her.
Description
Low Noise model is optional. See model description for details.
FAQ
Comments (25)
Rather hard to get rid of glasses even with negative prompting.
Yep. She wears glasses. :) like some of us, she just has to wear them all the time and they're part of how she looks.
That said, you will have better luck dropping glasses if you only use high nose model and you don't use lightx2v. Negative prompt only works with cfg higher than 1.0
@darkroast175696 You said "Negative prompt only works with cfg higher than 1.0" and I think I'm seeing that my negative prompts are completely ignored! What do you recommend I modify from your (already awesome and fast light-based workflows) so that negative prompt is respected while still being fast & beautiful?
@BigJuggsAI Yes, as I said, the negative prompt only works with a cfg over 1. As far as I know, that's a hard rule so you have to choose. You either live without a negative prompt in order to get faster generation, or you need the negative prompt so your renders are slower.
You can use a higher cfg with the lightning loras, but if you boost the cfg over 1.0, they will be much slower. It's still faster than doing a full render without lightning since you can use fewer steps. So that's just a choice you have to make. There's no secret for working around it.
@darkroast175696 Thanks for sharing that hard-earned experience as I would have not know otherwise. Odd to me that awesome lighting would completely ignore the negative prompt but your workflows have great prompt adherence on the positive. (I assume that a little cheat like setting cfg to 1.1 or 1.5 would not be beneficial?)
@BigJuggsAI I think part of the point of the lightning loras is to get better prompt adherence even when cfg is 1.0. If you don't use lightning, you can still set cfg to 1.0 and wan (and any other renderer like flux or sdxl) will ignore the negative prompt. It's just what you get with 1.0, and a 1.0 render is always much faster per frame than any higher value.
@darkroast175696 Oh I think I understand... two two concepts are orthogonal and not related to one another... however if people have lighting installed then they are interested in top speed and for that cfg 1.0 is the speed king?
Q: What would you say is a cfg value that would restore my negative prompt? What performance loss in percent should I expect over cfg 1.0? (If it's just a few percent then I'll put back your recommended setting so negative works again but if it kills performance then I won't)
What an outstanding Lora! Your best work! Congrats!!
Thanks very much!
how to keep the face in western style when other loras are applied? can increase the size of breasts without adding other loras? Can no glasses?
If you want Mio but with a different face, your best option is to use only the high nose model at about 0.75. You can try even lower strength and see what works for you. It's hard to get rid of the glasses. You can use a prompt to help, like "she has a bare face without glasses" but that only works most of the time in my experiments. You can also use a low noise Lora like a character model or one of the Instagram girl models or something and that will also influence the face. You don't have to use the low noise Lora at full strength. Again, just try out different settings until you get the look you want. Make low res test videos so it won't take you days just to try things out.
@darkroast175696 i love this lora very much. But the breasts size of my works is smaller than your demo, if I added your bursting out lora, girl's face, and girl's leg become thick. I hope you can create more asian style beauties with extreme huge breasts lora, Thanks... Excellent works
@superwai777974 You might be using slightly different settings from me in your comfyui. There are so many different ways to set up a wan 2.2 render with different lightning loras and samplers and schedulers and so on. Here are all the settings I used for the sample video where she stands up from the water wearing the red bikini with white dots:
Steps: 5 / 5
Size: 640 x 960
CFG: 1.0
SD3 Shift value: 5
Sage Attention: yes
Sampler: res_multistep
Scheduler: sgm_uniform
High Noise LoRA (all at 1.00 strength):
Mio high noise
BoobPhysics v7
Wan_2_2_I2V_A14B_HIGH_lightx2v_MoE_distill_lora_rank_64_bf16.safetensors
Low Noise LoRA:
Mio low noise - 1.00
BoobPhysics v7 - 1.00
lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16.safetensors - 1.25
Seed: 633910569828320
Prompt:
close-up shot, natural light, edge lighting, soft lighting.
mio.
a pretty young japanese woman with short black hair and round wire-rimmed glasses is emerging from the water in a swimming pool. her hair is wet and plastered to her skin. she stands up out of the water suddenly drawing attention to her body. the water cascades off her wet hair and her body. her head is tilted back and her hands are sweeping her wet hair back away from her face
wearing a crimson bikini with a polka dot print.
front view camera angle.
the camera slowly pans and pushes in, capturing the woman from a dramatic and sexy angle.
@darkroast175696 So for (awesome) Mio to be Mio we 1) need at least 0.75 in high noise and the remaining 0.25 and all of low noise is 'somewhat disposable' as far as her amazing body shape and breasts?? (What does adding the rest bring? Asian skin texture and face?)
(I'm trying to create a character that has the same glorious breasts but can be tuned to be other races like African American or Latina) Thoughs?
@BigJuggsAI If you only use the high noise model, and only use it at a strength between 0.5 and 0.75, you will get Mio's body shape, but you have more freedom to customize other details. It isn't 100% reliable, but it's pretty good. You will just have to experiment to get the look you want.
@darkroast175696 Thanks for sharing your A+ hard-earned experience! You're the best! Happy to return the favor with Wan 2.6 experience! (Hint: It's AWESOME!) PM me at BigJuggsAI @ DA
Have you ever released your wildcards?
@GayLizardSpy No, not yet. That's on my list of things to do. I feel like I should clean them up a bit and organize them some first, but I do plan on releasing.
Edited to say: I'd be surprised to see you using them, though, since your prompts and videos are already so off-the-wall and creative.
@darkroast175696 Gotcha, I'm compiling my first wildcards now. Maybe my videos will become more "sane", lol.
@GayLizardSpy No, please don't make them more sane :)
@GayLizardSpy Here they are!
https://civitai.com/models/2430843/darkroast-wildcard-pack
@darkroast175696 Awesome! Thanks, can't wait to try them out!
@GayLizardSpy I hope they give you a good starting point for making some of your own stuff. It's really nice to be able to write a big randomized prompt or a series of several prompts that will probably take several attempts to get right and then just set it running when you go to bed. Then you get up in the morning and see what Wan got you for Christmas.
Also thanks for the buzz tip!