What is this

Look alive! 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 smirking facial expression. Smirking is an asymmetric smile which can convey a range of nuanced emotions: a smirk can be smug, it can be mischievous, or it can even be flirtatious. It's more than just the mouth; the whole face, and especially the eyes, need to work together to form the expression.

How to use it

The trigger word is "smirking".

Start with 0.7 weight.

Download both smirkLeft and smirkRight. Try one of them, and if it doesn't work, then you try the other one (for some images it's more natural to smirk to the left, and vice versa).

If your image doesn't have a closeup face composition, don't use this in the initial generation, but instead inpaint the face area. Use sufficient padding when inpainting (see the sample images for what composition is expected).

How was it created

I trained left-side-smirk and right-side-smirk separately, because it was easier to teach these concepts to Stable Diffusion separately than together (results were better this way). This separation also allows more fine grained control for the user. Both left and right were trained with the same settings, same images, and same captions. The images were just flipped horizontally.

Why was it created

Why did I choose to train the smirking expression specifically? It's an interesting and nuanced expression that the SD1.5 model can partially understand - just not well enough to work with consistent characters without a LoRA. So it's perfectly situated between "too hard, can't make it work" and "too easy, no need for LoRA". I'll also note that no smirking LoRAs have been published before this one. For more background on my research into facial expressions on consistent characters, see my article at https://civarchive.com/articles/4563/facial-expression-flexibility-with-consistent-characters