Wan2.2 - French Kiss (I2V/T2V) - CivArchive (CivitAI Archive)

Wan2.2 - French Kiss (I2V/T2V) - High Noise 1.0v

NSFW

The purpose of this LoRA is to provide a better french kissing generation then the base model. The base model does kissing and some degree of french kissing but its not good. We want a little bit more tongue and not kiss peck!

Features

Longer french kissing sessions
Optional 'tongue play' prompt
Compatible with both I2V & T2V Wan2.2 Models. Primarily geared towards I2V.
Trained on close-up and full body.

Tips

Low Noise - try between 0.5 to 1 strength (I2V). Lower strength if you are experiencing face changes. Closer to 1 for (T2V).
This model technically works without the Low Noise model (I2V) because the base model is trained on kissing to some degree.
May need to bring the faces closer otherwise could do some air french kissing. "the man leans towards the woman" type of thing.
Try prompting like this for longer french kissing scenes:

(AT 0-1s) the woman leans towards the woman and begins french kissing her

(AT 2-4s) the woman continues to french kiss the woman.

Slow Motion Issues? https://www.reddit.com/r/StableDiffusion/comments/1mr0sep/how_to_fix_slow_motion_in_wan_22/

Prompts

the woman french kisses the man

the woman are tongue playing as they french kiss each other

Other

Trained on Wan2.2 T2V base model. 59 high quality images and 33 videos that were handpicked and farmed raised to perfection using musubi-tuner. This LoRA works with both I2V and T2V.

Shoutout to all the LoRA creators out there doing the insanely tedious work of tagging, editing and running training sessions

Description

FAQ

Comments (21)

MysticalOneNov 28, 2025· 7 reactions

CivitAI

This is outstanding! I have been looking for a better kissing lora as a starter for my video sequences. Still testing it for lora weights, but so far the results are amazing. It also works seamlessly with other loras if you have the characters doing other things simultaneously! Please keep iterating on this one, it's amazing!

MysticalOneNov 28, 2025

In my testing, no matter what lora strength for this you never lose consistency in the characters' faces. Which is great if you are forced to balance multiple loras, you can dial this down to what you need if another lora needs more strength. Great job! Also works perfectly with lightning loras, I didn't see any distortion of motion or fidelity.

Kierkegaard420

Author

Nov 28, 2025

@MysticalOne I'm so glad to hear! I'm redoing my old LoRAs (adding more video/images, too) to resolve the face change issues.

skullzy77Nov 28, 2025

idk what I'm doing wrong but it doesn't work at all for me. Will have to play some more after work tonight

MysticalOneNov 28, 2025

@skullzy77 When you say it isn't working, what is happening? Be sure you are prompting like @Kierkegaard420 mentions in the description, those triggers are important.

skullzy77Nov 28, 2025

@MysticalOne Yup, I was using their prompt. It would just mash their mouths together and if it did show tongues they were melted together in a blob. Most of them were really slo-mo also. I tried different strengths, tried without the low lora and tried without the high lightx2 lora. Nothing I did changed anything.

Kierkegaard420

Author

Nov 28, 2025

@skullzy77 There's likely an issue with your step count. Increase steps and make sure you are correctly switching from high to low noise following Wan2.2's suggested process. My cover examples are using lighting at 12 (total) step count which is 3x more then the recommended use for lighting. The model needs more details with motion (high noise) and more details in general (low noise). At least that's been my experience. Might be able to go below 12, I'm sure someone can comment.

stbeerystudios662Nov 28, 2025

CivitAI

Does a very good job. I can't get it to have the characters sustain a kiss. Tried lots of prompt variants without success. I really like the model though. Thanks

Kierkegaard420

Author

Nov 28, 2025

@stbeerystudios662 Try different seeds and make sure you are following the tips above. You need 'continue' the action otherwise the model will change to the exit part of the kiss.

Vyxen808Nov 28, 2025

CivitAI

Works fine, but my problem is that it makes the kisses look slow. I don’t know what the issue is, but it doesn’t happen when I use other LoRAs

Kierkegaard420

Author

Nov 28, 2025· 1 reaction

@Vyxen808 The dataset does inherently have 'slower' kissing action since there is not really a lot of data out there with fast french kissing. It should be at least as fast as my posted examples. If you have speed issues slower then the cover examples, then its a step issue or issue with handling the transition from high to slow.

erdelman73267Nov 29, 2025· 2 reactions

just use cfg high and low at 3.5 and use SEKO high and low lighting loras at str 1, and do at least 15 to 25 steps, Euler/simple. It should give u more action and normal movement. (use WAN MOE K sampler)

Kierkegaard420

Author

Nov 29, 2025· 1 reaction

@erdelman73267 Thank you for your suggestion.

erdelman73267Nov 29, 2025· 1 reaction

@Kierkegaard420 and almost forget he should use WAN MOE K sampler. I just uploaded a video that was generated such way it should show up soon.

Kierkegaard420

Author

Nov 29, 2025· 1 reaction

@erdelman73267 Yeah, https://github.com/stduhpf/ComfyUI-WanMoeKSampler it awesome for handling the correct transition from high to low steps.

LuxariaNov 28, 2025

CivitAI

This is made more for I2V than T2V, you should change the description.

Kierkegaard420

Author

Nov 28, 2025· 1 reaction

@Luxaria I'm sorry but what is the issue? This model is literally trained on the T2V base model and my sampling that was run during training, using T2V, shows its working fine.

Kierkegaard420

Author

Nov 28, 2025

@Luxaria Here's a T2V on the first try: https://civitai.com/posts/24740051 I'm not great with T2V so it could probably be refined more but it seems to be working for me. I've updated the notes to indicate that this model is more specifically geared towards I2V but it definitely does work with T2V.

LuxariaNov 28, 2025

@Kierkegaard420 on my gens this Lora for T2V has no effect at all, it just gen the Wan2.2 base kissing type. The tests I did on Wan2.2 14B Q8 + Q8 Text Encoder, CFG 3.5 and 20steps.

Kierkegaard420

Author

Nov 29, 2025

@Luxaria I'm sorry to hear.

asimangunsong998109Feb 4, 2026

CivitAI

great work! what would be your recommended setup for combo with lighx2v lora?

LORA

Wan Video 2.2 I2V-A14B

by Kierkegaard420

Download (Beta) View on CivitAI

Details

Downloads

12,561

Platform

CivitAI

Platform Status

Available

Created

11/27/2025

Updated

6/11/2026

Deleted

Files

WAN2.2-FrenchKiss_HighNoise.safetensors

Size:

292.63 MB

SHA256:

bba38b08ea78b661d20cef533f9cf9e371556ec70d1846031ac7f6c6c1490dc3