As always please use adetailer. I'm still working on my eye lora for FATE. I hope you all enjoy the model. I generated most at 1024 x 1536. Please leave a positive review if you liked it.
Progress notes
Thankfully, this dataset was largely already done. Back in December I had a bold idea of creating a lora for Cu's various forms. It ended in disaster. I still need to edit a few images, particularly for full body and portrait folders. Fucking kill me. He actually has an ear clasp. I need to redo a ton of the pictures now. He doesn't have nipples and the hair covers where they would go. I wonder how generating this will work. He has an index ring in the anime screenshots. Back to redoing the editing. I'm not sure how the gloves are going to go. He wears a long sleeve black shirt, but its under his cloak. I don't want to confuse the AI with all that layering since its the hood, the gold trim, the white trim, the black shirt, and the cloak is covering the sleeves. So I opted for a glove and fingerless glove tag. Let's see how this goes. Test 1.
Test 1 ended in failure. He overfit way too fast. I'm reducing LR and TE and trying again. I've reduced repeats as well. Starting training for test 2.
Test 2 ended in failure. Wrong patterns on clothes. Raising LR/TE. Training for test 3 now.
Test 3 failed. The LR isn't high enough. I'm returning to original LR with more training images and reduced repeats. It is still less steps than the original so hopefully no overfit. I am curious what would happen if I put in the eye lora I'm working on as a dataset in this. Sadly, that eye lora isn't ready yet. I'm still building the up and editing the dataset. Getting ready to start training for test 4.
Test 4 ended in failure. The armlet was overexaggerated on the single_armlet but the gold trim was learned. I've removed a few images of the armlet but will that make a difference? Starting training for test 5.
Test 5 suffers the same issue. Adding back some removed images and reducing the LR, then starting training for test 6.
I don't know whether this test was successful or just luck. I do think I would run it again at a slightly lower LR but for now I will test the epoch to make sure its actually good. It's not consistent enough. Lowering LR and starting training for test 7.
Not good enough. Retesting.
Test 7 failed. I'm going to add more index ring photos in the closeup subset and then retrain.
Test 8 failed. I'm going to edit more screenshots of the rings since that seems to be my biggest issue. But someone, please tell me, why they show his detached sleeves as actual sleeves that cutoff at the hands in some frames but then in other frames they are actual sleeveless gloves that go around the fingers!? I hate inconsistency. T_T It's more work for me. They do this a few times actually. In another frame his glove is like triangular and wraps around one finger. BUT I will lower the LR again. I actually prefer using lower LR. Let's see how this goes. >_> It's probably another night of lora training. Starting training for test 9 now.
Failed. Adjusting LR and starting testing 10.
Failed. Adjusting LR and starting test 11.
Failed. Adjusting LR and starting test 12. I think due to how weak the index_ring tag is as a danbooru tag the model may have trouble pulling it. I've raised the TE to 50% of LR to help compensate. Could this have been the issue the whole time? I am unsure. I hate high TE due to how much it overfits. Starting training for test 12.
Test 12 is successful. I don't know whether it was the TE or LR. I want to learn more towards the TE. I also ended up changing the resolution so that was probably a factor as well.