AI_LoRA_3종 test - 100-hours-a-week/16-Hot6-wiki GitHub Wiki

상위 문서로 이동 : AI Wiki

LoRA 소개

OTT LoRA

• 데스크 셋업 이미지 366장으로 학습한 우리 서비스 특화 LoRA
• Kohya_ss_LoRA_trainer_XL로 학습 Link참조
• Base Model: Fluently-XL-v2, SDXL 파인튜닝 모델로 이미지 생성 테스트시 조금 더 나은 성능을 보여주어 해당 모델로 설정 Link참조

3D_Officer

• CivitAI에 제공된 SDXL 기반 LoRA, Link참조

Dalle-3-xl-v2

• SDXL기반 LoRA, Link참조

LoRA Settings

• dalle_only: Dalle(1.0)
• ott_only: ott(1.0)
• 3d_only: 3d(1.0)
• dalle_ott: dalle(0.7), ott(0.5)
• dalle_3d: dalle(0.7), 3d(0.5)
• ott_3d: ott(0.7), 3d(0.5)
• mix: dalle(0.6), ott(0.4), 3d(0.4)

Test 1

• Prompt: "A clean, modern desk setup in a white-toned minimalist room with an ultrawide monitor and a neat mechanical keyboard, photorealistic, ultra high resolution."
• Negative Prompt: "blurry, low quality, noisy, distorted, deformed, bad proportions, text, watermark, messy, cluttered background, cartoon, anime, painting, sketch"

	dalle_only	ott_only	3d_only	dalle_ott	dalle_3d	ott_3d	mix
image
VRAM	약 8.72GB	약 8.72GB	약 8.72GB	약 8.72GB	약 8.72GB	약 8.72GB	약 8.72GB

Test 2

• Prompt: "A clean modern desk setup with a dark wooden desk, a large monitor, a white mini PC, potted plants, a wireless ergonomic mouse, a mechanical keyboard with colorful keycaps on a felt mat, natural daylight, cozy and minimalistic."
• Negative Prompt: "blurry, low quality, noisy, distorted, deformed, bad proportions, text, watermark, messy, cluttered background, cartoon, anime, painting, sketch"

	dalle_only	ott_only	3d_only	dalle_ott	dalle_3d	ott_3d	mix
image
VRAM	약 8.72GB	약 8.72GB	약 8.72GB	약 8.72GB	약 8.72GB	약 8.72GB	약 8.72GB