Ryo Nameoka from Blue Lock! Please use adetailer. If you have any requests please let me know. If you enjoy the model please leave some positive reviews!
Progress notes
So the issue is, as you can guess, the logo. Tests 1 and 2 finished and the issue is sustained.
I'm going to train heavily on the uniform subset for round 3 and maybe slightly increase TE so it binds better? We'll see. Maybe I should try add a logo tag first.
His hair is all over the place. In some shots his spikes are left. In some they are on the right. Sometimes his spikes cross each other too. It's a nightmare.
I don't know what test I'm on. It's his logo that is causing issues. I hate logos so much.
The recent test failed. Raising network dim/alpha and conv dim/alpha to help.
Not looking good so far. I wonder if the issue is its conflating the DaDaDa with logo as opposed to the much smaller logo. My theory now is---I need to make copies of all the shirts that have the DaDaDa and the logo, remove the DaDaDa text and keep only the logo for the tag so it learns to associate it better.
Let's hope this works. I'm changing the dim/alpha and conv dim/alpha back to 32/16
Well, it improved a little. But not enough. Ok, made a new folder for tight crops. Increased dim/alpha to 64/32 along with conv. Reduced my LR as well. Let's hope this round goes better!
There was some success with this one...HOWEVER...TE is way too high. also concept bleeding from dark green stripes turning whole jersey green. i'm debating lower LR still....
It improved but it's not there.
I'm trying raised LR now since I do think my folder is better.
The lower LR/TE did help with 96/48 dim/alpha. However, its not strong enough. Or the crops are harming the generation of portrait shots. Adjusting repeats on crops and portraits and raising LR/TE
I do think I am finally finished with him. However, I don't think it learned his hair as well as it should have. On the other hand, his hair had different spikes depending on the angle so it was always going to be hard to learn. I'm ready to move on though.
Nevermind. The hair is going to bother me too much. I'm going to try again.
The hair is still not good enough. I am going back to 64/32. I did not experiment enough with these settings. The 96/48 is overfitting too easily.
Still no luck. I'm going to try adding more images.
I've decided to remake his dataset from scratch. In the editing process now.
It's 3:28AM. I've finally finished editing. I photoshopped the eyes in every single image in hopes to improve their quality. I added more tight crops as well as images of upper body and full body with the logo. Let us hope it works this time.
Face was rough, background was weird. Retraining at half LR/TE.