Readme
This is a mild nsfw lora. Trained specifically to depict a character in various poses and enable consistent character control. It contains no sex scene datadset.
Works well with I2V & T2V.
Supports character loras without affecting the face. Base dataset has a tendency to generate an average normalization dataset type of face. If you want to create a unique looking face describe face characteristics and face expression
Prompting
zerozeroam - you can do just fine without the lora trigger keyword. Keyword enables dataset trained at extreme close-up
High LORA works best with low weight (0.1 - 0.3)
Low LORA works best with high weight (0.7-1.0)
I used the following prompt structure:
<lora trigger keywords>
<description of a character, clothing, setting, mood>
<keywords>
<action>
protip 1: blur the I2V image and add "focus on person"
protip 2: use joycaption to generate a prompt to better represent your idea
Keywords
(choose one or mix from each category)
young adult, adult, mature
obscured, tiny pussy, small pussy, medium pussy, large pussy,
skinny, thin, fit, medium, large, very chubby
standing, leaning, laying, sitting, crouching, on all four,
relaxed, standing straight, standing flexing, leaning forward, leaning backward, laying on back, laying on stomach, laying on a side, sitting on a floor, sitting on a chair, flexing, relaxing,
one arm up, arms up, both hands up, arms down, arms crossed, legs down, legs up, legs crossed,
arms together, arms in t-pose, legs together, legs to the side
in front, slightly to the side, side, from behind,
low angle, from below, frog perspective, high angle,, from above, drone perspective,
in detail, close up, focused on a character, from a distance,
correct, upside down, laying on side
looking at camera, looking away, looking slightly over the should, looking over the shoulder, looking down, looking up
looking at facing up, facing to the side, facing down,
wearing clothes, half naked, completely naked
naked outfit, casual formal, dressed up,
no makeup, gentle makeup, heavy makeup
breasts under clothes, partialy visible, bra, half exposed, naked upper body,
flat chest, tiny breasts, medium breasts, very large breasts,
happy, joyful, neutral face, worried, teasing, curious, sad, excited, playful,
white skin, black skin,
Technical
Sample videos generated with only lighting lora on a base wan 2.2. model
Rank 64
Trained on 480x720 dimensions dataset
Description
FAQ
Comments (9)
Nice! So finally this will look natural...
https://civitai.red/images/126832886
Thanks for the LoRA. Could you provide information on how to create the prompt using joycaption?
You can try using joycaption from huggingspace eg https://huggingface.co/spaces/fancyfeast/joy-caption-beta-one
Give it an input image and prompt to precisely describe an image eg
Write a very long detailed description for this image. Do NOT mention the image's resolution. Do NOT mention any text that is in the image. Do not mention the mood/feeling/etc of the image. You MUST use at least dozen of keywords from the prompt to write a very long caption: <here is your list of copy pasted keywords>
Writing the prompt is a topic on it's own, often trial an error, you must speak like a model itself. A good prompt for wan consists of senteces that work, and avoid bs that doesn't. Sentence eg "The weather is nice" is so generic it does more damage than it helps.
And how to describe action? Don't write "Person throws a rock", write "Flexing their muscles person reaches to the back with their arm and in a split section, in quick and dynamic motion moves their hand in an arc forward throwing a solid rock." or something like that. It's not the easiest thing to do, but the more you observe captioning models like joycaption do it, the more you learn.
Hope it helps, even if just a little
epic awesome, the only problem is that you need gpt to describe the action, but rest is epic cool.
I don't really understand what this LoRA is supposed to do at all. Helps motion from changing from one pose to another?
"Trained specifically to depict a character in various poses and enable consistent character control"
My problem is that most character/concept loras present human - well woman - as either super sexualized plastic dolls or uninspiring blob of bodyfat, caused by bad bad prompting used for the training process of the model. In my loras I want to create something that more resembles subtle, mild nsfw that has some aesthetics to it.
V1 tries to achieve just that while being flexible - use it for static images, use it for animation. Tbh I'm not doing anything new, just filling the gap that was left unnoticed. Wan 2.2 is a powerful model but it needs some direction or guidance. Most available loras fail to comply to it's rules.
I think v1 definitely fails at a few things which I'm going to address in v2.
@_00am When are you planning to release version 2?
I don't get it, could this be useful in SVI workflow to keep character during face turns?
I didnt understand this lora until I had a challenge to make a character to take certain pose, which most of the loras have but not quite from independent action or modifying with strong prompting, so, I remembered this lora and tested it and yup it solved it!!
Awesome work!! 👏🏻👏🏻👏🏻👏🏻