Hi all! I hope you have fun with this one. Please share your creations here if you can, and in any case, thanks for checking it out. Cheers!
Trigger words: licking teeth, licking lips, biting lip (full prompts and ComfyUI workflows are provided in my example images)
(Trained with Diffusion-Pipe on Windows via WSL 2: https://github.com/tdrussell/diffusion-pipe/)
Description
Note:
Trained on images only with T2I in mind, though it works a little with video (only tested with T2V) just because Wan is smart enough to kinda figure out what to do once the pose is forced
Full prompts and ComfyUI workflows are provided in my example images/videos
For whatever reason the Wan version really struggles with the "biting lip" pose; you can see from my examples that it's possible, but it was a constant challenge with plenty of seed surfing, so if you run into issues there, it's definitely the model's fault and not yours (good luck)
Cheers!
FAQ
Comments (4)
I can't seem to get the lip biting to work for the z image version. the mouth looks all deformed and the picture itself is blurry. any fix?
@TypicalLoraLover Sorry to hear that it's been giving you trouble!
It's hard to diagnose without more knowledge of where/how you're generating the images. Are you using ComfyUI by chance?
If so, you could try saving one of the Z-Image lip-biting examples I've uploaded, then use the ComfyUI menu's "File > Open" option and load that PNG image, which should load in the example workflow (I only used native ComfyUI nodes, so there shouldn't be any missing nodes for you to deal with unless your ComfyUI is an older version).
Also, a known issue with Z-Image LoRAs right now is that image quality tends to degrade dramatically when using multiple LoRAs together (this is due to the Base model not being released yet, so LoRA training at the moment can only be done with some clever hackery that doesn't play well with multiple stacked LoRAs).
The main workaround to avoid this is to make sure that the total combined strength between all of the LoRAs you're using is 1.0, though of course in some cases this reduction in strength could nullify the affect of the LoRA(s) you're using. There's reportedly also a custom node that helps selectively apply only the important blocks of each LoRA to help mitigate this problem, but I haven't tried it yet.
Good luck and cheers!
@redcreamsoda yes im using comfyui. i tried your workflow and sadly i still cant get the image to look good. mouth is still deformed and image is still blurry. i am only using one lora (your lora) and i am not using natural language to prompt... hopefully this info helps a bit
Any chance of a Zi Base version too?