I ended up using different resolutions when generating. You'll find a combination of 832x1216, 1216x1792, 1344x1920, 1344x1024, 1536x1024. Changing the resolution was the only way I could get everything to generate properly in full body photos due to the latent space. Ajeel does have a widow's peak but we never see him without his bandana. I edited his appearance so I could generate him without the bandana. But I like the bandana more. I did try to generate him without nipples because Fairy Tail characters don't have nipples. As always, please use adetailer. I hope you guys like this model!
Progress notes:
Dataset prep is rough. He has a widow's peak which you can see in some angles. But there are no photos of him without a headband.
Just finished test 1. I stayed up all night and spent 11 hours working on the dataset. I didn't like these images. I'll work on adding more portraits to the dataset as well as fixing tattoo. I need to remove the black_tattoo tag. The issue that came up was the tattoo was appearing on both arms.
Starting training for test 2 now. I've created portrait folder and tattoo folder.
The tattoo/guild mark isn't accurate enough to my standards when the character is facing forward. Re-editing dataset and then testing again.
It's still not capturing the detail accurately. Adjusting batch size and dimensions to help. Test 4.
Not accurate enough. Adjusting batch size and dimensions to help. Test 5.
It gets better towards the end of the training but it's still not there. I assume the LR is starting too high since the meaningful learning is happening closer to when its low. Adjusting and starting over again for test 6.
I failed again. It wasn't bad but the headband isn't being learned properly still, or at least passably. Starting test 7.
Test 7 failed. It was an issue with angles or the batch size. I'm adjusting repeats and switching over to L2 for detail emphasis.
Test 8 failed. I don't like the quality of it. Switching back to Huber.
I feel better about the tattoo. However, the headband is still an issue. Originally I thought it would learn the headband from the images alone. However the yellow_headband tag is fighting against the design on the headband. I've created the ajeel_headband tag and am now making a dedicated subset for learning the headband as well. Starting test 9 after.
The closeups and upper body are pretty good but once again the full body eludes me. I'm retraining with an extra tag for tattoo and will generate full body at higher resolution. While in some ways this feels like defeat to me as long as the lora is capable of generating it as close to accurate it will be a success. Some may call my obsession with details insanity. Starting test 10.
I think I have found some success but I am on the fence about it. I am using a new resolution, 1216x1792. It does really well. And you don't need adetailer. The only concern I have with it is people not being able to use resolution that high.