<h1 id="wan-2.2-kiss-lora-v0.2">WAN 2.2 Kiss LoRA v0.2</h1>This is the second version of my WAN 2.2 Kiss LoRA.After completing my first training experiment, I gained some useful experience and learned a lot from the results. I am now releasing Kiss LoRA v0.2, with a larger dataset, more detailed captions, and additional training steps.<h2 id="what's-new-in-v0.2">What’s New in v0.2</h2>Compared with the previous version, this release includes several major improvements:<h3 id="1.-larger-training-dataset">1. Larger Training Dataset</h3>The training dataset has been expanded from 90 videos to 180 videos.Every training sample is a complete 5-second, 81-frame video clip, rather than a selection of individual frames.<h3 id="2.-more-detailed-captioning">2. More Detailed Captioning</h3>For the first version, I used only one caption for every training video:<code>kiss</code>I did this mainly because I was being lazy, and I could not find a reliable method for automatically captioning kissing videos.For v0.2, I manually reviewed and categorized the training clips more carefully. Different kissing behaviors now use different trigger words, allowing the LoRA to produce a wider range of actions.<h3 id="3.-more-training-steps">3. More Training Steps</h3>The number of training steps has been increased:<ul><li>v0.1: 1,000 steps</li><li>v0.2: 1,500 steps</li></ul><h2 id="trigger-words">Trigger Words</h2>The following kissing behaviors were included in the training dataset:<h3 id="kiss"><code>kiss</code></h3>Standard kissing, with a stronger focus on lip-to-lip contact.This is the most stable and reliable trigger word.<h3 id="passionate-kiss"><code>passionate kiss</code></h3>Produces more intense and energetic kissing movements.This prompt generally creates more passionate body language and stronger interaction between the characters.<h3 id="tongue-kiss"><code>tongue kiss</code></h3>Focuses more on visible tongue interaction during kissing.<h3 id="wet-kiss"><code>wet kiss</code></h3>Intended to produce wetter lips and more visible saliva.However, this trigger word is currently not very reliable.<h3 id="lick-kiss"><code>lick kiss</code></h3>Intended for kissing while licking the other person’s lips.The results are still inconsistent.<h3 id="tongue-lick"><code>tongue lick</code></h3>Another experimental trigger for tongue-licking behavior.It does not perform consistently in the current version.<h3 id="tongue-suck"><code>tongue suck</code></h3>Intended to create tongue-sucking actions.This trigger word does not appear to work very well yet, possibly because there were not enough matching training samples.<h3 id="swap-spit"><code>swap spit</code></h3>Intended to produce visible saliva exchange.This trigger is also unstable and may require several generations before producing a good result. In some cases, the model may incorrectly interpret saliva as other types of fluid, so expect occasional artifacts or inaccurate generations.<h2 id="recommended-base-model">Recommended Base Model</h2>I do not recommend using the official WAN 2.2 model with this LoRA.In my experience, tongue movements generated by the official model often look unnatural or artificial.The base model that currently works best with this LoRA is:WAN 2.2 Smooth Mix I2V V1This is also the model I personally use for testing.Smooth Mix already has some native kissing capabilities, but its kissing actions tend to be repetitive and lack variety. However, its motion quality is excellent, and it can make kissing scenes feel much more dynamic and passionate.For now, I believe it is the most compatible base model for this LoRA.<h2 id="about-me">About Me</h2>I’m a kiss lover.I created this LoRA purely as a personal passion project because I enjoy kissing scenes and like exploring the many different ways a kiss can be presented.Kissing is far more varied and interesting than it may first appear. Different movements, emotions, rhythms, and interactions can create completely different results.When I realized that very few people were creating LoRAs specifically focused on kissing behavior, I decided to make one myself.I am still a beginner when it comes to training video LoRAs, and I also have a busy full-time job. I can only work on this project during my spare time.If you are also a kiss lover, feel free to share your feedback, suggestions, test results, or training ideas. You are also welcome to participate in the project so that we can experiment, improve the LoRA, and share our results together.<h2 id="future-plans">Future Plans</h2>All current versions were trained using a 512 × 512 bucket resolution.I am not sure whether training at a higher resolution, such as 768 × 768, would produce better details or more accurate interactions, but it may be worth testing in a future version.I also plan to experiment with additional behaviors, including:<ul><li>Face licking</li><li>More varied tongue movements</li><li>Different kissing positions</li><li>More expressive and emotional kissing styles</li><li>Improved saliva interaction</li><li>More stable close-up kissing</li></ul>However, collecting and preparing high-quality video datasets is extremely time-consuming. I am also very selective about the quality of the training videos, which makes dataset preparation even more difficult.Because of this, the next version will probably not be released very soon.But there will definitely be another version.Thank you for downloading, testing, and supporting this project. Please share your results and let me know which trigger words work best for you.<hr /><h2 id="my-first-wan-2.2-lora">My First Wan 2.2 LoRA</h2>This is my first attempt at training a video LoRA.It is not perfect, but it works—and for me, that already makes it a meaningful first step. The LoRA was trained on Wan 2.2, which I still find to be a reliable foundation for character-focused video generation.I also experimented with LTX 2.3, but in my experience, it still struggles with character consistency and identity preservation. Facial features can become unstable or distorted, especially during larger movements or longer generations. Because of this, I decided to continue training with Wan 2.2 for now.Hopefully, more powerful and consistent open video models will arrive in the future. I am looking forward to seeing how the ecosystem develops—and to training better versions as I gain more experience.<h3 id="trigger-words">Trigger Words</h3>No trigger word is required.<h3 id="prompting">Prompting</h3>The LoRA is designed around scenes featuring two people.And the prompting just : kiss .Example prompt:<code>two people kiss.</code><code>They are kissing.</code>For better results, try adding:<ul><li>Detailed descriptions of both characters</li><li>Clothing, hairstyle, and facial features</li><li>Camera movement and shot composition</li><li>Lighting and environment</li><li>The specific interaction or action</li><li>Mood, pacing, and visual style</li></ul>Since this is my first LoRA, results may vary depending on the base model, resolution, sampler, LoRA strength, and generation workflow.Thank you for trying it. Feedback, example videos, and workflow suggestions are always welcome.

cxy_kiss

WAN 2.2 Kiss LoRA v0.2This is the second version of my WAN 2.2 Kiss LoRA.After completing my first training experiment, I gained some useful experience and learned a lot from the results. I am now releasing Kiss LoRA v0.2, with a larger dataset, more detailed captions, and additional training steps.What’s New in v0.2Compared with the previous version, this release includes several major improvements:1. Larger Training DatasetThe training dataset has been expanded from 90 videos to 180 videos.Every training sample is a complete 5-second, 81-frame video clip, rather than a selection of individual frames.2. More Detailed CaptioningFor the first version, I used only one caption for every training video:<code>kiss</code>I did this mainly because I was being lazy, and I could not find a reliable method for automatically captioning kissing videos.For v0.2, I manually reviewed and categorized the training clips more carefully. Different kissing behaviors now use different trigger words, allowing the LoRA to produce a wider range of actions.3. More Training StepsThe number of training steps has been increased:<ul><li>v0.1: 1,000 steps</li><li>v0.2: 1,500 steps</li></ul>Trigger WordsThe following kissing behaviors were included in the training dataset:<code>kiss</code>Standard kissing, with a stronger focus on lip-to-lip contact.This is the most stable and reliable trigger word.<code>passionate kiss</code>Produces more intense and energetic kissing movements.This prompt generally creates more passionate body language and stronger interaction between the characters.<code>tongue kiss</code>Focuses more on visible tongue interaction during kissing.<code>wet kiss</code>Intended to produce wetter lips and more visible saliva.However, this trigger word is currently not very reliable.<code>lick kiss</code>Intended for kissing while licking the other person’s lips.The results are still inconsistent.<code>tongue lick</code>Another experimental trigger for tongue-licking behavior.It does not perform consistently in the current version.<code>tongue suck</code>Intended to create tongue-sucking actions.This trigger word does not appear to work very well yet, possibly because there were not enough matching training samples.<code>swap spit</code>Intended to produce visible saliva exchange.This trigger is also unstable and may require several generations before producing a good result. In some cases, the model may incorrectly interpret saliva as other types of fluid, so expect occasional artifacts or inaccurate generations.Recommended Base ModelI do not recommend using the official WAN 2.2 model with this LoRA.In my experience, tongue movements generated by the official model often look unnatural or artificial.The base model that currently works best with this LoRA is:WAN 2.2 Smooth Mix I2V V1This is also the model I personally use for testing.Smooth Mix already has some native kissing capabilities, but its kissing actions tend to be repetitive and lack variety. However, its motion quality is excellent, and it can make kissing scenes feel much more dynamic and passionate.For now, I believe it is the most compatible base model for this LoRA.About MeI’m a kiss lover.I created this LoRA purely as a personal passion project because I enjoy kissing scenes and like exploring the many different ways a kiss can be presented.Kissing is far more varied and interesting than it may first appear. Different movements, emotions, rhythms, and interactions can create completely different results.When I realized that very few people were creating LoRAs specifically focused on kissing behavior, I decided to make one myself.I am still a beginner when it comes to training video LoRAs, and I also have a busy full-time job. I can only work on this project during my spare time.If you are also a kiss lover, feel free to share your feedback, suggestions, test results, or training ideas. You are also welcome to participate in the project so that we can experiment, improve the LoRA, and share our results together.Future PlansAll current versions were trained using a 512 × 512 bucket resolution.I am not sure whether training at a higher resolution, such as 768 × 768, would produce better details or more accurate interactions, but it may be worth testing in a future version.I also plan to experiment with additional behaviors, including:<ul><li>Face licking</li><li>More varied tongue movements</li><li>Different kissing positions</li><li>More expressive and emotional kissing styles</li><li>Improved saliva interaction</li><li>More stable close-up kissing</li></ul>However, collecting and preparing high-quality video datasets is extremely time-consuming. I am also very selective about the quality of the training videos, which makes dataset preparation even more difficult.Because of this, the next version will probably not be released very soon.But there will definitely be another version.Thank you for downloading, testing, and supporting this project. Please share your results and let me know which trigger words work best for you.

Files with SHA256: 75c068e48fae77d4a3e5d4fe23e7267e71f340852c02161af639685bbe1bd029