Our favorite American-Lebanese reporter and media personality, Mia Khalifa!
Honestly, this one was hell to finetune. Due to her time in the spotlight being rather short, it was surprisingly tedious to gather a workable dataset for her. Had to rerun the process up to 4 times and do a lot of testing.
The embed has been trained on SD 1.5 and works well on most 1.5 based models. Due to limitations in the dataset, there are difficulties getting a wide variety in backgrounds and clothing. Using the Embedding at a lower weight like Koh_MiaKhalifa:0.85 or using inpainting helps a lot.
If you enjoy my creations and want to support me or have a face you'd like to commission, consider supporting me on Ko-Fi or Buy Me A Coffee :)
Description
FAQ
Comments (22)
Any chance you could do Tori Black (Michelle Chapman)?
Saw your comment yesterday and here it is! Fabulous! ^o^
Can you do Denise Milani please?
What settings did you use for training the embedding?
I'm trying to make my own and it just doesn't seem to be going well.
Hey, really the settings are only part of the process. You have to take care to gather a good quality dataset too, anywhere from at least 20 to 75 pictures. How often do you save embeddings in between? For Mia, I had to repeat the process multiple times, saving embeddings every 50 steps.
There's a lot of randomness involved in all of this as well, given the nature of AI/machine learning and the insane variety of the human face.
@SDKoh i used 23 pictures and saved every 50 steps. The problem is probably the learning rate. i only have 8 gb so i'm also limited by that
@KingPungy actually, in my experience the learning rate can stay at default (0.005). I train on a 6gb videocard so that shouldn't be a bottleneck for training. Have you turned off any VAE in your settings? Training with a VAE can give bad results
hit or miss
I agree with you man. The potential is definitely there but the consistency is lacking. Was a very tedious proces to even get it where it is, might redo it once I've done a new face or two.
I guess they never miss huh?
@Osangni You got a boyfriend, I bet he doesn't kiss ya
@SDKoh Skies is referencing a meme, not critiquing the model (or maybe both, I dunno)
how many images have you used for the dataset?
Hi Caine! For Mia I used about 40 pictures, but depending on the face of the subject I go anywhere between 25 and 100. A bigger dataset for Mia would have been good since her face is rather 'extreme' (wide proportions, very wide mouth, rather big nose, very strong eyebrows) which makes it prone to being overtrained.
@SDKoh thanks for the advice, can I ask you also how many batch size and gradient accumulation have you used? I'm trying to replicate a good pt like this
Hi, what model did you use on your showcase images? Mia looks fire ,thanks
Hi, the first image is made in RealisticVision V1.3, the second and fourth in Protogen 5.8. Both are great models, you should definitely check them out. For some reason it escapes me how I made the third, I think it is Deliberate or DreamMaker.
Just wanted to say she has done interviews after her time in the adult world. Perhaps that might help your dataset if you make any revisions! Check out her twitter time line if you want to.
Thanks mate, I'll take this into consideration if I get to finetune her face in the future. I looked around through my files and it turns out I did rip her twitter account with RipMe, not sure why I didn't look take pictures from it for the dataset
can you do hande ercel plz?
Oh wow, what a beautiful face! I'll definitely look into that, thanks for the suggestion mate
her ig is full of photos, rip it and use those for ref, idk why i decided to look one up for her but glad youre working on it
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.