260613 Anima version
Unexpectedly, the Anima version was much more difficult to train than SDXL. As a result, it became a very unstable and difficult-to-use Lora.
In Anima, the tendency to express character details is so strong that no matter how much I supplemented the dataset or increased the rank values, I could not avoid the AT field being drawn behind the character. Even a slight increase in the number of tags representing the character causes the AT field to be pushed into the background.
However, if there are a few solutions:
1) Increase the weight of trigger words (up to around 2-3) and decrease the weight of the character (around 0.5-0.8).
2) (I do not know the reason) but the success rate tends to increase when author tags are added and other style Lora elements are added.
3) When editing with the inpaint function, it tends to be rendered semi-transparently without any particular issues.
Anyway, it is a very lacking Lora, but I have uploaded it for now to conduct various tests. If anyone knows of a good solution, please let me know.
This is an incomplete Lora. Initially, I wanted to apply the AT-field between the two characters, as well as the AT-field extending toward the sky, but it was beyond my capabilities.
Placing a translucent object in front of the character was surprisingly difficult. Illustrious kept drawing the character first and pinning the AT-field to the background. I tried many times, but it still takes a lot of luck to properly draw the AT-field in front of the character. If anyone knows a good solution, please let me know.
Trigger : atfield to front, energy barrier, transparent barrier, atfield in front of the charanter, character_partially_occluded, octagonal pattern,
There are a few imperfect but effective ways to improve the success rate.
1) Try slightly increasing the weight of the "atfield in front of the character, character_partially_occluded" tag.
2) Place the AT-field tag before the character description tag.
3) Because of 2), it tends to be a bit more unstable on noob-ai-based checkpoints.
4) Euler a sampler is more stable than DPM.