This lora allows one to generate a fairly good likeness of the celebrity/actress Alycia Debnam Carey. If I were to personally rate the accuracy I'd put it at least at 95%. Happy generating.

Hello, I’m back! This time with SDXL training. To be honest, this has been a project for me, to get Lora’s as accurate as possible. I always aim for 100%, but if I did that, then the Lora would probably be over trained ;).

I’m going to be real here. I really don’t know what the heck I’m doing when it comes to training a celebrity lora. I delve into the internet realm, from reddit to other random websites to gather advice (read advice not asked) on achieving accuracy on lora training. I get annoyed, use some dummy parameters and keep tweaking until I find my sweet spot.

At least for this Lora, (of Alycia Debnam Carey), the single best piece of advice I was able to find was, selective image samples. My database of ADC was probably twice what I actually used. I actually had some previous loras made up, but I was not at all satisfied with them. The subject (Alycia Debnam Carey) was lost quite often, with 1 in 5 or 6 generated images being of the subject. I did something wrong. Images were not the biggest issue in actuality, my dataset text got corrupted. It was missing the initializing prompt :o… for half of the dataset. So yeah, understandable on why I wasn’t getting what I would like.

However, it does not change the fact that the actual image dataset did corrupt the subject matter. What I am about to say is by no means a scientific fact, made by repeated observation and testing backed up by fellow constituent data, it is simply a matter of my case with repeated observation and testing. The cherry picking of images did help, considering I aimed for training image data according to ratio, 1:1, 2:3, and ??16:9?? not sure about the last ratio, just know it was landscape. So all of my images were cropped for the ratio. In my opinion, at least when it came to landscape, a major issue in image generation. I know what I want and over estimated what I wanted.

I aimed for upper body, headshot, half body, and full body. Issue with that, is the images available to me were in most cases high quality, but subject became lost because of cropping and lowering the quality as a result, especially in regards to full body. By cutting out that single bit of training data, it drastically improved generated subject quality. My primary focus on generations are no greater than half body for subjects anyhow. So where the wild idea came from that I needed full body images for training is beyond me, since it holds no water in my primary use case, save for a few ideas here and there.

So with that the cutting of full body from my dataset, it was a significant reduction. Cutting landscape training, reduced my dataset, then the cherry picking of viable images for SDXL reduced my dataset. It brought the grand total of training images to 98… almost hit a hundred there ;). With that, 5 repeats for 20 epochs. A grand total of 9800 steps. Not to shabby. And a critical reduction of training time by a significant margin.

I’m still touting a 6800xt for Lora training. Would love a NVIDIA card just to play around with but I needs that Nvidia money and I’s ain’ts got its. Total time for the Lora to be trained was about 6hrs. Like I said, a critical reduction from the 14 hours I had before. I could not even achieve the full 20 epochs with the dataset before. I got to maybe 15 or 16 before my maximum step count kicked in. (My personal maximum is 14000. I’ve trained other SDXL loras before running a 20hr training and after 14k, to much burn. 14K offers the best accuracy without burn for a large dataset, well large for AMD…)

Well, now that I’ve completed this one, I’ve got revisions to make to others. Some cherry picking, some deleting and the whole nine yards. Not an exact science, but determination will get you somewhere eventually.

As this model is based on a real life individual, an actress, Alycia Debnam Carey, I do not give my permission to use this Lora for the production of NSFW content. If it would be inappropriate at work or to show to your mother, then it shouldn’t be made or posted. The loras I invest in are admittedly of individuals I find myself attracted to, but in no manner, shape, or form, would I abuse the usage of their visage for lewd content. Therefore you, the user, allowed to borrow my work, shall not either.

Be a human, and keep in mind, these are real people, with real feelings, real motivations, fears, wants, needs, etc. If you do that, then hey, have fun. Nothing like generating a celebrity wearing a sombero in a goofy fashion.