This LoRA should play well with other LoRAs without needing to mess with strength much.
All videos contain metadata required to generate them.
Here's a sample Prompt (the // are comments, so you can leave those in):
//Simple description of the scene
Two gorgeous women in wedding gowns passionately kissing eachother.
//Describe the characters
They are both thin and gorgeous.
The woman on the right has her hand on neck of the woman on the left. She has her long blonde wavy hair adorned with white flowers. She is wearing a long white lace wedding dress.
The woman on the left has her long red hair in a fancy bun. She is wearing a short silk wedding dress.
//Describe the background/environment
The background a white wedding trellis with flowers. They are in a garden with lush greenery.
//Style of the video
The video has a delightful bright vibe. It is a realistic video.
//Describe the camera angle
long-shot. Realistic.Description
WAN version of the passionate kissing LoRA
Trained at 12,500 steps at 2e-5
FAQ
Comments (21)
Which WAN model is this for? 1.3 or 14b....
14b
also works for male and female i2v. Thank you!
Something I am just realizing is that when you have two model types (in this case Wan and Huny) uploaded to the same page like this one, you can't search by both model types, only Wan. I assume it's whatever the latest version is, but now I am wondering how much stuff I can not find because it's mixed into other models, or marked wrong. Anyways, great model though, I downloaded a while ago but had to find it again for the prompt.
This is a very good LoRA. It blends with character LoRA. Did you crop out the face of your dataset? or included it?
Faces can't be cropped out when kissing. I typically captioned:
- face shape
- eyebrow color and size
- eye color and detail
- eyelash length
- makeup
- lip size
- hair color, style, and length
- skin tone
- any tattoos or piercings
and any other unique features. Prior to finalizing captioning these, I typically try to generate a 1 frame video of the face without kissing to see if the model can generate something very close to that face with my captions. If my captions get really close, that face will not train into the captioning.
@ComfyTinker Anything we're captioning will be trained as a unique feature, won't it? So you mean that if we are able to caption the face feature, we can still include the face without further overlapping with the face of Character's LoRA?
@YesPleaseProceed I actually wrote an article on captioning: https://civitai.com/articles/11942/training-a-wan-or-hunyuan-lora-the-right-way
A lot of guides and people don't understand what's happening behind the scenes, so hopefully that helps you out.
@YesPleaseProceed In short, no, captioning unique features and facial components don't train the feature to the term "passionate kissing", that train them to the terms like "bushy eyebrows" or "long eyelashes", so yes my LoRA will affect eyelashes and eyebrows differently than the base model if you prompt them, but if you don't prompt them they will be unaffected.
I linked you the wrong thing! Sorry! I updated the link to the article https://civitai.com/articles/11942/training-a-wan-or-hunyuan-lora-the-right-way
Any chance to get this for the 1.3B model ?
For some reason you are only one of two people on here i found that even have any 1.3 loras.
(Thanks for that from an AMD pleb with a cheap(ish) card btw)
I attempted to train this for the 1.3B model, but it's super inconsistent. I can try again, but I don't want to deliver something that isn't very functional.
nice for risen
May seem like a stupid question, but is this T2V? ..or I2V?
Actually not a stupid question, the author is stupid for not adding this simple info.
@Learning2025 It's both. There was no option for separate T2V and I2V when this was created, as everything was only trained on T2V when WAN first came out, but T2V works for both. Quick to call someone stupid for this.
@ComfyTinker Thank you for taking the time to clear that up.
Let's not try to criticize our creators and try to help them along. You simply could have just suggested to the Creator that he had that information So that we would know.
Especially if they are taking time out of their day to create things that we download on the platform for free. So be more respectful.
does this work with the wan2.2 14b with the 4 step lora? If it does, I cannot get it to work much at all.
It works with I2V 14b with 4 step LORA. Just connect it to high noise model.
Well I'm glad I found this Now I don't have to spend a lot of money on those templates that don't work all that well and that you really can't prompt in some cases. So this actually works really good at about 1.5 -2.5 When I'm using the Wan2.2 14B model (Hooked up after the SD3 and before Sampler)
"(Kiss passionately)The girl quickly turns to the man and then she starts giving him a strong passionate French kiss with tongue, then wrapping her arms around his heck to pull him in close((Passionate Kissing))."
Details
Files
WAN_Passionate_Kissing_v1.safetensors
Mirrors
WAN_Passionate_Kissing_v1.safetensors
1474712_WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
137_WAN_Passionate_Kissing.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
wanPassionateKissing.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
WAN_Passionate_Kissing_v1.safetensors
Available On (2 platforms)
Same model published on other platforms. May have additional downloads or version variants.