Update: Also try now in XL!
Part of my facial expressions series. All of these LoRAs are designed to mesh well with each other and have low impact on your image framing.
An expression that it's almost impossible to coax from SD otherwise, clenched teeth and neck from exertion or disgust.
This was a really hard one and I'm still not totally thrilled with the outcome. SD kind of seems to just legitimately not know that the bottom row of teeth exist. Fascinating really. If you need it to work harder, I recommend adding emphasis to "grimacing" in the prompt, not the LoRA weight. It works up to 1.6 or so for me pretty well.
Would love to see how it works for folks, so please share any feedback or outputs. It's definitely harder to work with animated models, so I gave a specific prompt for it below.
Recommended negatives: tongue
Starter prompts:
<subject>, grimacing expression
animated:
1girl, grimacing expression, teeth
Description
FAQ
Comments (14)
finally, thanks !
Loving these! Any plans for an XL version down the pipeline?
Sadly not any time soon. I train locally and XL performance is still a little prohibitive, both in terms of training / memory and in terms of how long images take to generate for testing and evaluation.
Even if I had the hardware, to be 100% honest, finding good, diverse, representative native-quantity 1024x1024 source images would be a little tricky, a ton of the training corpus data I collect is barely 512x512 and I'm skeptical of the feedback loops from any upscaling / AI image generation techniques (I only ever use real, native, ground truth in training).
Will update if that changes, though! SDXL is definitely very cool, and the expanded encoder is clearly capable of representing a ton more complexity in prompts, so I hope things streamline down at some point (or that 24gb memory cards become a lot more accessible)
@thegipper Was just asking, I'm happy enough using these on 1.5! They're amazing! They're the best realistic expression LoRas I've come across! And yea, I know XL is different and tricky - but ya never know until you ask! I seriously use these all the time. Smashing!
@Picky_GoHickey Ha! For sure, and sorry for the exposition dump, always curious to learn more about folks who've made the jump and relative differences given how little I've worked with it.
Thanks for the feedback and let know if you have any thoughts or suggestions!
@thegipper Are you speaking of the jump to XL? I use both, but due to the speed and limited number (compared to 1.5) of checkpoints and LoRas for XL, I primarily use 1.5 I admit, though - XL is 10x more responsive to prompts, and, given the right model and prompt/neg combo, easily able to outshine 1.5 image quality wise hands down. And due to XL issues like the one's you mentioned regarding training XL LoRas, it's harder to tell about flexibility. It's very flexible when responding to prompt changes alone, but a lot of LoRas are less flexible due to being trained on fewer images or runs due to resources or time. If things ever catch up to make XL as easily as 1.5 - like if magically a crazy converter existed that allowed all 1.5 loras to work on xl, then I have no doubt xl is more flexible. But that's not the reality....yet. And suggestions? You bet I have some! It's nice to see your contributions for high quality non-anime expressive loras. I'd go for an "eyes bulging" one - like a comical eyes popping out of head kind of thing, There's plenty of those for anime that don't translate to realistic, sadly. And there's 565 models for aheago or tongue-out, but why are they all the same? The tongue can bend many ways! Or an eye rolling / rolled one? Again, always seems to be anime. Or...sorry for my spitballing here - you could have an 'extreme' series of realistic yet pushed to the extreme reactions? If that makes the separation of normal and extreme easier on you. If that makes sense...I'm rambling. Anyways, that's enough jibberish for now! Honestly, unrelated to anything above, the greatest mystery to me is how precisely 2 checkpoints are able to handle realism so smoothly where other models fall apart or turn jagged/broken - and how to apply that one attribute to other checkpoints or as a lora. Seriously, play around with Stablegram and Beautiful Realistic Asian V6. How do they do that? Can always turn to those for crazy extreme images/loras that look terrible on most other models. It's crazy!
Excellent!
We could do with an "attitude"/"sinister"/"evil" smile expression LoRa, your expressions are really good, could you try to make that?
Thanks!
That's a really fun idea! I'll plug around and see if there's good data to be found, might take a while to come up with clean samples, but now I'm curious if SD can pick up on the nuances.
@thegipper Fantastic! Thanks, look forward to trying it when it is done!
Flux DESPERATELY needs all of these expressions. You should be able to train with all your old training images here on CivitAI. I have been getting amazing results.
Good stuff! Flux please.
I've been working on it, but with somewhat mixed results. Most of the training I've done seems to work great with the base model, but much more poorly when mixed with subject LoRA's of individuals (which the SD1.5 LoRA's do really well). I can't quite figure out if that's an issue with the subject LoRA's being overtrained, or an issue with how I've trained the expression LoRA's, but it seems consistent with both people LoRA's I've trained and the most popular ones here on Civitai.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.