This is a complementary file to the LoRA posted earlier, you can use it as standalone too but it's recommended to use both at the same time for improved resemblance. If using together, tone down the LoRA weight somewhere around 0.8-1.0
Description
FAQ
Comments (7)
What ckpt was this trained on? This and the Piper models (TI & LoRAs) both seem to have an issue with jawlines that are much too narrow and chins that are extremely triangular in shape. It's almost as if they were trained on a ckpt that is biased towards asian faces.
I've tried several photo-real models trying to eliminate this problem, but it seems to be baked into the LoRAs and TIs.
From the nose and up, the resemblence is excellent. However, from the mouth and down, it throws off resemblence to a fairly large degree.
All my realistic models are trained on base SD 1.5, I would not use anything else for training. You can adjust face shape quite a lot by playing with weights of both LoRA and TI (that are supposed to be used together ideally). But if there is a problem, it can be only in the training set - I didn't have anything better than quite blurry 768x576 screenshots (probably taken from VHS transfer) so it is what it is. If you can get me DVD quality screenshots from first season, I can try to re-train it.
Actually yes, you might be on to something, but: 1. it's a matter of preference (she also has mouth slightly open on almost all examples what naturally makes chin appear longer and narrower), 2. it's controllable by adjusting TI weight. Here's an example with TI weights changed by 0.1 from full to 0.7 weight (while LoRA is at full weight): https://ibb.co/44SWmqT If you want just different face shape, this should work fine.
@dolirama126 I don't know what I am doing wrong. I am using Photon and all of your settings, I'm using your advice by playing with the lora and ti strength settings and am getting a result that barely looks like her. In fact, I've had the same thing with many of your recent models. I just cannot get them to resemble the intended character. I don't know why. I've tried different checkpoints too.
@spoofmeister492 You should post some examples with metadata, then I might be able to tell where the problem is. Are you using both LoRA and TI together?
@dolirama126 Yes, I am using both. I have also tried them seperately. Ok, I will post some examples when I get a chance.
@dolirama126 Dolirama, there is now on Civitai an option to talk to other members
on-line live.
If you are in need of very good HQ photos of a certain actress, you can contact me.
I have access to a huge selection of Stars (classic stars and new) for you to make your
models. You do not need to use screencaps from old VHS sources. Only in certain
cases , specially for rarely seen actress or actors, which there are not many pictures
to use for model training and you have to get screenshots from films.
But since I used to be a celebrity faker for over 25 years, I have collected a huge amount
of photos. I can get screenshots from films and then re-work all the details in images to
a High-quality using Photoshop and other programs I have for doing A.I. models.
You can also contact me by pm in the forum Planetsuzy if you need my help.