<h3 id="heading-720">What is this</h3><a target="_blank" rel="ugc" href="https://civarchive.com/tag/look%20alive!">Look alive!</a> is my series of facial expression LoRAs that are intended to work together with consistent character LoRAs and embeddings. This particular LoRA is for the puckered lips facial expression.<h3 id="heading-7">Intended use cases</h3><ul><li>kissing face gesture (dataset contained only images of this type)</li><li>pov kissing generations</li><li>duckface expressions</li><li>blowing a kiss expression</li></ul><h3 id="heading-7">How to use it</h3>First try with simply the trigger word, "puckered lips".Experiment with the weight and use the smallest possible weight that gives you the result you want without too many artifacts.If you have a lot of stuff in the prompt, other embeddings and LoRAs, etc. then you might find it difficult to get the expression without cranking the weight too artifact-orily high. In such a case I recommend you experiment with prompting:Positive prompt: ((puckered lips)), mouth closed, flirting, kiss face, kissing lips, kissing mouth, kissing, (blowing kiss: 0.5)Negative prompt: ((teeth, open mouth, mouth open)), smiling, soulless gaze, blank expression, dead eye stare, doll, figurine, statueIf you have an issue where the mouth is not fully closed and increasing the weight of the LoRA/stronger prompting leads to bad looking teeth, the solution to this issue is to increase the weight and make the prompting stronger until you get the mouth fully closed. When the mouth is fully closed, no teeth will show. (My dataset does not have any images with teeth, because it only contains images of strong expressions where the mouth is fully closed.)If your image doesn't have a closeup face composition, you may want to consider inpainting the expression after the initial generation. Use sufficient padding when inpainting (see the sample images for what composition is expected).<h3 id="heading-7">Comparison to alternatives</h3>I created this LoRA because SD1.5 is not capable of generating this expression when used with consistent characters (even though it is capable of generating this expression without consistent characters). For more background into my research on facial expressions, you can read this article: <a target="_blank" rel="ugc" href="https://civarchive.com/articles/4563/facial-expression-flexibility-with-consistent-characters">https://civarchive.com/articles/4563/facial-expression-flexibility-with-consistent-characters</a>To understand how this LoRA works relative to your alternative options, I recommend looking at the comparison grid posted as sample images:<ul><li>The first image in the grid (1girl, no lora) was supposed to demonstrate that SD1.5 is capable of doing the puckered lips expression when you're not using a consistent character. However, the seed was unlucky and it doesn't show a good expression.</li><li>If you look at the first column (no lora), you will see that none of the consistent characters I used were able to create the puckered lips expression without the help of additional LoRA.</li><li>If you look at the second column - that's this LoRA - you will see that this LoRA helps create the puckered lips expression on all of the characters used, though it also introduces minor artifacts and alterations to the image.</li><li>If you look at the remaining columns, those represent other LoRAs on Civitai that have been released for the same purpose. I put in some effort to provide a fair comparison by experimenting with the weights and by reading the instructions on each model page and adjusting the prompt to match the instructions as best as possible (though I was not following the instructions 100% because I had to be able to make it work with A1111 XYZ grid script).</li><li>The third column represents the <a target="_blank" rel="ugc" href="https://civarchive.com/models/11998/emotion-puppeteer-lora">"Emotions puppeteer" LoRA</a>. I had trouble getting the expression with this LoRA so I cranked the weight pretty high. As you can see there's more artifacts compared to my LoRA and the expressions are still weaker.</li><li>The fourth column represents the <a target="_blank" rel="ugc" href="https://civarchive.com/models/308591/expressions">"Expressions V3" LoRA</a>. Here I had to lower the weight a little bit because at 0.7 weight the generations were extremely cartoonized. Even with the lowered weight the generations are still a bit cartoonish, and the expressions are not as strong as they are with my LoRA. You can also see various lighting artifacts that look like the training dataset of that LoRA may have had many images in oversaturated studio lighting conditions.</li><li>The last column represents the <a target="_blank" rel="ugc" href="https://civarchive.com/models/108931/puckered-lips-helper">"Puckered Lips Helper" LoRA</a>. You can see that it makes better "puckered lips expressions" compared to my LoRA, but it introduces many alterations in the generations. Most notably, all the characters seem to have excessive makeup, enlarged lips, and the shape of their jaw is altered. I suspect that that LoRA has been trained with a dataset where the same person (with large lips and excessive makeup) has appeared in too many images.</li></ul>

Files with SHA256: 43b78a8e7ba7d7f34d3cca4c290b8f387dfcc4dff51c71ac6b4e78c2a8992764