Bite that lip!
Krea 2 v1 - 7.1.26
This is everything I ever hoped, y'all. Really loving Krea2. However, it is a bit weird on attention. In my testing, I found that it really latches on to every detail, sometimes a bit too much. It seems to mix with certain other loras flawlessly, and others it just crashes. So you'll have to experiment some.
Krea 2 v1 Tips:
Keep your descriptions of mouth activity super limited. Use the "she bites her lip" and that's it. Maybe another generally expressive word like seductive or flirty. But anything else related to the mouth or lips will absolutely go bonkers. So no "smirk, pout, smile, parted lips, open mouth" kind of stuff. It gets confused.
Use lower strength. I think it is maybe a little overbaked. But I liked the results of the 3000 step version at 0.55 strength alot more than I liked the 1600 step version at 1.0 strength.
Wan 2.2 I2V v1
Definitely happier with this run than the last one. Still recommend highest resolution you can manage. Also, I had really terrible results using SmoothMix checkpoint. Not sure why. Best results I have seen are using the base wan. Test videos use Q6_K.gguf models. Also, I recommend messing with the lightning lora weight. I got better stability using 1-1.5 weight on lightning lora, or 2-2.5 if using other loras for sex acts for more speed/movement.
Wan 2.2 I2V v0.1
I'm seeing better results at higher resolution. So 720x1280 seems to be the best. Lower resolutions have sometimes worked, but only if the face is large and the primary focus of the image. My guess is that in the compressed latent space, the lips are just too small at lower resolution for the model to focus on, much less for the Lora to affect. So if you need small resolution, make sure the lips are big in the image reference.
I have a version of the core workflow I use here:
https://civarchive.com/models/2076123/wils-style-workflow
Description
This is the one! You guys, this is the model that I think finally gets it. It is maybe a bit overtrained. The earlier checkpoints just didn't do quite as well. But it seems to work pretty well at lower strength. So like 0.55-0.75. Full strength works too, it just starts to pick up a more baked look and it affects facial features.
FAQ
Comments (12)
hey! I've been having so much fun recently with these lip bite loras you created! this might be my 4th comment lol, but would you be able to do a ltx 2.3 version for the lip bite lora? been wanting to get lip biting to work on ltx 2.3 but no luck...
I'm open to trying it. My biggest issue with that one is that I can't find an LTX workflow that will generate good results for me like I get with WAN 2.2. I'm sure I'm just doing something wrong, nobody else seems to have problems. I just haven't gotten over the hump of figuring LTX out. Building a lora takes not just the data and training run, but testing. If I can't get quality videos off the model without the lora, then gens with the lora will just be worse.
Do you have a workflow you use for LTX?
@darkwil i have a workflow yes!!! its a seed hunter workflow that i found somewhere on civitai! ill send my version of it!
@darkwil https://drive.google.com/file/d/17_XRINNK3kphFEG5mDY6vl9yMRUh5Hoi/view?usp=sharing here you go! if you can't get this workflow to work for you, the default ltx 2.3 should work too!
@TypicalLoraLover Thanks for sending. It is basically a more complicated version of what I had tried before. The core of the workflow like samplers and models is all the same. The omni lora helped some, so that was new, but my results are still artifact heavy and super poor motion. I suspect there's something else in my local comfy abomination of an environment that is messing it up. Like I've noticed Z-Image can't handle sage attention. And I've had some questions about the new comfyui dynamic vram setting. Maybe LTX doesn't like one of those? Dunno. I may try a different attention and see if that helps. And it is also heat of the summer now, so it may be a bit before I want to run a 500W heater in my office for hours. The krea 2 version snuck in during a cool spell. But... I dunno. We will see what happens, my friend. I'll keep tinkering.
@TypicalLoraLover Oh, also, I ended up trying something different with the krea2 lora where the caption was just the concept instead of a detailed caption and it worked really well. I may give that a go on klein at some point, see if that fixes the mangled training I was getting. You still really want that or is krea the new kid in town for you too?
@darkwil no problem! take your time! also you can run ltx 2.3 with low vram with wan2gp! I run it ok pinokio and I use it all the time for ltx 2.3 but if theres a certain lora that requires to use a specific sampler, i use comfyui, but i mostly use wang2p for ltx 2.3. hopefully this helps a little bit for ya! it's always another option if you wanna try to run the lip bite lora on wan2gp as well!
@darkwil krea seems to be my favorite at this point instead of klein. i can't deal with the problems that flux 2 klein has with arms and legs and hands and fingers for me...
@TypicalLoraLover So I tried a couple runs on LTX 2.3. My gpu just can't swing it so I tried in runpod. Either LTX doesn't like my dataset or I'm training wrong because... the checkpoints suck balls (not lips, unfortunately). I have one more trick I'll try on it, but not being able to experiment locally with it is tough.
@darkwil no problem! just do what you can! if it doesn't work, then don't worry about it!
Fantastic lora. Can you please do the licking lips with tongue one, as well?
Thanks for the comment! Lip lick? Maybe? I can look into it. I had to fight pretty hard with various models to put together my current lip bite dataset. I think getting a genuine lip lick would be just about as tricky, but I will try it. No guarantees though ;)