Previously on https://civitai.com/articles/25094/lora-tests-lora-locon-loha-lokr-glora I compared different LoRa types and concluded that LoKr is perhaps the best, and I most likely trained too few steps.
So here is LoKr again but with longer training.
LoKr with dimension and convolution 32, alpha 16, factor 0.
Trained for 18000 steps (16h on RTX 3070).
Resulting models are in archive "LoKr (long)" here: https://civitai.com/models/2343120?modelVersionId=2635649
Results
The model had noticeable progress until about 10k steps. After that it stabilized and did not benefit much from additional training.
It is hard to say which one of them is the best so with principle of "more is better" I choose 18000 steps as the winner.
Sometimes at certain steps the resulting checkpoint is somewhat bad for no apparent reason. Here we see recurring anomalies at certain checkpoints.
At 1700 steps the model tends to add a duplicate extra persons into the image.
At 6800 steps one specific tag is broken. It is possible, perhaps even likely, that other checkpoints also have broken tags. But finding them is practically impossible. The nightmare of untestably broken models.
Test images
Lets start with the same prompts as previous test.
nanami ao gets good results at 5100 steps already and doesn't seem to improve much after that.
kurumaki zakuro gets her dress details right at 6800 steps. Her attitude seems to improve a bit with more steps.
isone kotoha and nanami ao at step 11900 learns that there should be an orange line in the shirt.
All three of them gave cursed results at 17000 steps. Sometimes the result of certain step is just bad in some way and it is often hard to notice it. But at least here it is obvious.
The training duration was doubled from the previous test and we still haven't fried the model.
Sometimes long training causes the model to output images almost identical to what was used for training. Lets see if the model is overfitted like that. Here is an image from training set.
Here is output when the captions of that image were used as prompt:
Not too similar so probably not overfitted.
Lets try applying the style to some other characters not from this show.
Huh? What happened in steps 11900-15300? Lets try with different seed.
That is better. RNG can give weird results. Lets just ignore that.
Interesting how a red emblem forms and it then morphs into a pocket. In training data rin azuma has that red emblem in her school uniform while isone kotoha has chest pocket.
Uh what happened at 6800 steps? It is completely ruined. Later it gets better again. I tried more seeds and the 6800 step model is always broken with this prompt.
Also at 17000 steps we again have a second Frieren in the image.
After a bit of testing it turns out that "beach towel" is broken on that checkpoint. If I remove that from the prompt the output comes out just fine. It is weird how something like that can randomly break during training.
How about two unrelated characters?
Seems fine.
Do note that at 17000 steps we again have a second kishi touka on a poster in background.