Anime style mix (monogatari+evangelion)

This lora combines both Evangelion and Monogatari styles in single lora. This hopefully allows it to better learn that the styles are separate from characters, and to generalize a general "anime" style.

Useful captions:

• "evatv", "reb1", "reb2", "reb3", and "reb3_1" for various Evangelion styles.

• "monogatari" for general monogatari style

• "bakemonogatari", "hanamonogatari", "kabukimonogatari", "koimonogatari", "koyomimonogatari", "monogatari off monster", "nekomonogatari black", "nekomonogatari white", "nisemonogatari", "onimonogatari", "otorimonogatari", "owarimonogatari", "owarimonogatari 2", "tsukimonogatari" and "zoku owarimonogatari" for more specific monogatari styles.

• "anime screencap" for general anime style. I recommend having this at reduced strength like 0.5

When style is not specified the model tends to be a bit unpredictable with its style.

With just "anime screencap" it may randomly pick one of the above styles. I suspect that it needs larger number of anime sources to properly learn a "general" style.

If no "anime screencap" is used then it tends to produce very "not anime" style (see example image).

Training details:

Training data was basically just the previous Evangelion lora V2 and Monogatari lora V1 thrown together. Trained with 3443 images on Illustrious v0.1 for 7 epochs.

See the previous lora pages for more details about the training data