Thanks dude! It might be slightly undertrained though, since I only trained it until 350 steps.

@andyh28 Hey Andy! It really comes down to gut feeling/instict, developed through practice. And even with the experience of doing this dozens and dozens of times, I still get it wrong sometimes. I often find that people with softer yet characteristic faces train very fast. People with 'extreme' facial features (bigger than average nose, thicker than average eyebrows etc) overtrain very fast, in such cases I usually tend to use low learning rates and higher step counts.

Some Theory:
Another thing that influences total step count is the batch size and gradient accumulation. Say that we train a face with 20 images for 2000 steps, on BatchSize 1 and Grad.Acc 1. This means that every image gets seen by the AI about 100 times ((1x2000)/20 = 100).

If we do this same face with BatchSize 2 (and GradAcc still on 1), we would need only half the amount of steps for the AI to see each picture a hundred times, because (2 x 1000 steps) / 20 (dataset size) = 100x each image gets seen. While training a TI again yields unique results every time, this should create a TI that is roughly as strongly trained as 2000 steps on BatchSize 1.

Gradient accumulation in effect pretty much 'multiplies' the batchsize. So if we take a Batchsize of 2 and GradAcc of 2, we would only need about 500 steps to achieve a similarly trained TI. This is because 2 (batchSize) x 2 ( GradientAccumulation) x 500 (steps) result in 2000 pictures seen by the AI with a dataset quantity of 20, with each picture seen 100 times, ((2x2) x 500) / 20 = 100.

Higher batchsize pro's and cons
In my experience, using a high batchsize has a chance of smoothening imperfections or leaving out characteristic details on some faces. I suspect this is because when the AI looks at multiple pictures for example, it will try to reason what the 'average' face would be between those images. While smoothening imperfections is nice, smoothening out recognizable features is not. When I find this to be the case while training, I just restart training on a lower batchsize/gradient accumulation.

Randomness (RNG)
Sadly, randomness is a strong force when training AI. In the end, I am only suggesting or guiding Stable Diffusion's training, I can't strictly tell it what to do. Some faces will be harder to capture than others. Some faces will train in 350 steps, some in 3500. You might find yourself training with what you judge to be a good dataset, labeling and settings, but get bad results. Sometimes, rerunning the training with the exact same dataset and settings can yield very different results, so if you feel confident about your setup, give that a try.
This means that while you can develop habbits and workflows that yield a higher success rate, you'll likely never reach a 100% success rate when training. Sometimes, you'll have to edit the dataset, change the settings, or do nothing and just restart training with the exact same settings. It can be a bit frustrating at times, but randomness is a crucial part of the Generative component in Artificial Generative Intelligence such as Stable Diffusion.

@bojegi Thanks for the kind words Bojegi! I can definitely take a shot at Alison Brie as Annie Edison, but creating a dataset purely out of ripped screencaps from a 30fps 1080p show will be challenging, though I've managed to do something similar for my Terminator LORA :) Thanks for the suggestion.

@ElizaPottinger Damn, I wasn't aware some fellow creators had already tried their hand at her face. I don't really expect to succeed where others have failed, but I'll give it a try the coming days, who knows. One thing I got going for me is that I am blessed with the patience to drop to a 0.02 learning rate and test dozens upon dozens of TI versions :p

Hi Polar! I don't exclusively do celebs, at least not as a rule. Why, what kind of category of faces are you interested in? ^^

Hi Jecic, thanks for the suggestions! I'll have a look at their faces in the weekend, I'm a bit swamped these days. If you are willing to help streamline the process by providing datasets that's much appreciated, for sure! You can drop a link here, find me on the civit discord and DM me (SDKoh, like here on Civit) or message me through Ko-Fi or BuyMeACoffee, whichever works best for you! I have a slight preference for discord since I check it most often.

@jecic91716945 Hey that's great man! So, for datasets I generally need images that have a pretty clear face (it's okay if a bit of hair is in front of the face, but things such as masks, facepaint etc are not good). I tend to crop images around the face, usually with a bit of neck and/or shoulder visible :)

If you cant be bothered cropping, you can share the dataset with me and I will see to it. Thanks for taking the effort to source the images though, it's much appreciated :D

@jecic91716945 hey thanks a lot Jecic, could you keep them up until tomorrow night? I won't be home for another 24-ish hours.

I'll try to do at least one face this weekend since I haven't uploaded on civit in a few weeks :)

@jecic91716945 Just wanted to let you know i've downloaded the images and made a dataset of Elena Koshka. Saved me a lot of time with preselecting and cropping of images, thanks buddy!

I did one training run so far and in my opinion there is good likeness, but her lips are somewhat overtrained so I will retrain it with a lower learning rate to see if I can hit the sweetspot ^^ If second training session doesn't work out I'll just switch to one of the other 4 faces

@jecic91716945 hey dude, i have a version of Elena Koshka and Alexa Grace. I don't feel like they're spot on though. Do you have a discord so I can show you some of the ouput, for feedback?

Hi e06, I'm glad you enjoy my works :) A few months ago, somebody asked me the same here. I wrote this guide when I trained on a 2060 with 6gb VRAM which was very limiting, hence the low batchsize in my guide. However, most of the information about dataset creation and settings apply regardless of thd Batch Size.

I hope this helps for now, I have plans to write an updated article/guide on TI training in the coming weeks, but I haven't set myself a deadline for it