Lesbian make out session, trained on 52 videos from the same video series. Two models for I2V/T2V 14B..
The prompt: 2 women making out, 2 women tongue kissing, l3z_kiss.
I use strength 1.0 for the I2V model and the faces don't change..
Under suggested resources I list hearmeman’s workflows. I use his runpod templates for training and workflows - highly recommend. I also recommend the combine 2 character Lora flux workflow which can give you an amazing starting picture. I’ve been trying to train multi-character Loras with no success, so this is the next best thing.
Description
FAQ
Comments (22)
I'm organizing a solo picket in favor of eliminating triggers like l3z_kiss. Please stop using numbers and words that don't exist.
Why? it's a trigger on the style of the lora. It is an all encompassing, hopefully unique phrase that points to only the style. In some ways, it can confirm that you are absolutely prompting the training data, if for some reason your descriptions never match up. Ideally the training set will use descriptions that are expected, but the keyword makes sure there is no ambiguity. Is it because it's harder to pick only components of the style? Or simply the annoyance of having to type it out?
Perhaps I see the argument if we were talking about merging LoRas, as I believe it'd just merge that tag list into the new checkpoint -- meaning you could lose reference to the entire style. However, this should be the goal of proper tagging during LoRa training. All the weights should be the representation any clip should match up with, assuming it's trained well enough. I don't really know honestly. If there is a real argument, I'd like to hear it.
@makiaevelio543 In different builds for Stable Diffusion, there are default plugins that show which key phrases are used in LoRAs and allow you to add them to the prompt along with the LoRA. But in ComfyUI, working with LoRAs and triggers is implemented in a catastrophically terrible way. Every time I have to deal with specific triggers, I end up writing them into the LoRA's name. And people who name their LoRAs with numbers or typos force me to rewrite them from the LoRA's title. Sometimes a LoRA doesn’t even work with those words because you also need to include almost specific sentences, and I have to hunt down the LoRA’s page to copy the prompt.
If I don’t want to mix one LoRA with another, I just disable it. I don’t need these triggers.
@altolstyh386 right click the lora loader and view lora info, usualy the trigger tags are there
@lunarevan Okay, now I feel stupid.
Read a bit the good habits when training a LoRA before talking nonsense. Using triggers that do NOT exist in the database of the checkpoint used for training, avoids confusion of concepts by the AI being fine-tuned. Especially when the trained concept is something that the base model doesn't do well, or hasn't been trained on at all.
@lunarevan That's a new one for me, which loaders do this? For example, HunyuanVideoLoraLoader doesn't, as far as I can see.
@sarashinai I actually tried to get my trigger words in the right click “view Lora info” section but didn’t have the time to figure out how to make it happen. I use it all the time, certain Lora nodes don’t have that option but the native one does.
@SD_AI_2025 right - I thought kissing was a general action so I chose a dorky l3z to differentiate.
When it works, it works rather well, but about 50% of the times it seems to create strange blurs and lights around the area of the kiss, and even on the faces, to the point where the video is garbage. It happened with both strength model 1 and 0.7, I don't know if maybe I should use lower.
Are you using the skip node? I saw this when I was skipping 10 and teacache was turned up.
Also I use .6-.7 strength on this.
@jalenbrunson good call on the skip -- one of Wan's developer's recommended it even, but I even noticed it give bad results as low as -2
Under suggested resources I list hearmeman’s workflows. I use his runpod templates for training and workflows - highly recommend. I also recommend the combine 2 character Lora flux workflow which can give you an amazing starting picture. I’ve been trying to train multi-character Loras with no success, so this is the next best thing.
I don't see a combine 2 character Lora flux workflow on the site. Does it just create an image with 2 of the same char and then inpaint to change one to the other char?
@Kiefstorm Yes that's what it does, just makes it very easy. It uses flux and it's here: https://civitai.com/models/837857?modelVersionId=937395. I used it to make the sample image for this lora.
Well, well, well, it seems that today we are going to have fun.
its cool how this works better than image kissing lora's because video teaches the ai that kissing is a sequential process of sorts. looks great
Appreciate it! This was my second non-character Lora, hoping to improve on whatever I want to tackle next.
Do you mind sharing your training settings, specifically LR, Rank, resolution, length? I've not had luck retraining https://civitai.com/models/1152160 for Wan so far. Also I have 32 videos of women making out and tongue kissing, each 6 seconds in length and of various aspect ratios for the dataset for that model if you're interested in sharing data. Not publicly because it's not stuff I own it's from like PH, Xvideos, GKG series, etc but like in DM or something. Cheers.
Absolutely, sent you a DM.
also works for male and female i2v. Thank you!