top of page
20240318 人物一致性及风格一致性测试
ROLE
AIGC Engineer
DESCRIPTION
Generated and produced character and style consistency through LoRA.
PROCESS
Dataset Creation:
For SD15-LoRA, approximately 15 images are needed for stable output, though the more images you use, the more consistent the style will be. The final result becomes more stable with more images.
Model Creation:
Trigger Word: Flat illustration
Weight: 0.8
Substituting different macro models and weights may give more interesting results.
Test Results:
-
Stable output with consistent style.
-
Nearly consistent character output.
-
If more consistent characters are needed in the future, a separate model refinement for the character can be done. A simple model using 15 images takes about 2 hours to refine.
-
2024-03-29 Update:
Unexpected discovery: This LoRA, due to the use of large amounts of orange and green, can add an orange and green stylized effect to the images. This works not only in 2D styles but also in live-action styles.
bottom of page