Loading
Du hast einen ComfyUI-Workflow? Erstelle direkt im Browser eine verständliche Workflow-Dokumentation.Workflow verständlich machen.Jetzt ausprobieren ↗
This lora combines Evangelion, Monogatari, Yozakura Quartet, Konosuba and Amanchu styles in single lora. This hopefully allows it to better learn that the styles are separate from characters, and to generalize a...
Modellquelle
Quellenbeschreibung
This lora combines Evangelion, Monogatari, Yozakura Quartet, Konosuba and Amanchu 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:
"evangelion", "evatv", "reb1", "reb2", "reb3", and "reb4" for various Evangelion styles.
"monogatari", "bakemonogatari", "hanamonogatari", "kabukimonogatari", "koimonogatari", "koyomimonogatari", "monogatari off monster", "nekomonogatari black", "nekomonogatari white", "nisemonogatari", "onimonogatari", "otorimonogatari", "owarimonogatari", "owarimonogatari 2", "tsukimonogatari" and "zoku owarimonogatari" for various monogatari styles.
"yozq" for yozakura quartet.
"amanchu!" for amanchu!.
Quellen
1 QuelleModellartefakte
1 ArtefaktQuellenauszüge
1 AuszugSource context: 289 downloads · Type LoCon · Base model Illustrious
"kono subarashii" for konosuba.
"anime screencap" for general anime style.
If no "anime screencap" is used then it tends to produce very "not anime" style.
Multiple styles can be used simultaneously for varying results.
I forgot to caption End of Evangelion. So that one has just "evangelion". "kono subarashii" is already a bit biased towards megumin so perhaps a new unique caption for that style could have been good idea.
The results of such large scale lora are promising. It learned multiple styles and concepts appearing in them well but less frequently appearing characters were not so well learned. But the long training makes large loras inconvenient.
Training details:
Trained on 7416 anime screenshots from Evangelion, Monogatari, Yozakura Quartet, Konosuba and Amanchu. The model was trained for maybe 8 epochs (59328 steps). Training took so long that I had to save the state and resume training multiple times and I didn't keep track on how much each training run did.
Trained on Illustrious v1.0.
