티스토리 뷰

반응형

스테이블 디퓨젼 "작은 차이가 명품을 만듭니다" - Upscale 설정 값에 따른 이미지 변화

오늘은 Upscale by 숫자 셋팅 값에 따라 이미지가 어떻게 변하는지 여러 이미지를 통해서 실험해 보겠습니다.  

결론부터 말씀드리자면

1) Upscale by 값이 1.0에 가까울 수록 인물을 제외한 배경 prompt 에 대한 충실도가 높아짐

2) Upscale by 값이 2.0에 가까울 수록 얼굴이 디테일하게 묘사하면서 미인으로 바뀜.

     그러나 가끔 팔, 손가락 등의 왜곡이 일어남

3) Upscale by 중간값 정도에 본인 스타일의 여성이 나올 수 있음. (대표사진 처럼 ^^)

 

1. 적용 Prompt  및 setting값

A captivating portrait of a confident and stylish woman walking along the iconic Brooklyn Bridge in New York City. The bridge serves as a stunning backdrop, symbolizing connection and progress, while the woman embodies strength and determination, (wearing white turtleneck dress) (An extremely delicate and beautiful work), (masterpiece), (best quality),(beautiful), ((looking at viewer)), lovely lips, cinematic light, ultra-fine painting, delicate face, delicate figure, Fine collarbones, ultra detailed, highres, (realistic, photo-realistic:1.4), 8k, raw photo, physically-based rendering, professional color graded, 8K, F2.4, 35mm. ulzzang-6500-v1.1

Negative prompt: ng_deepnegative_v1_75t, (two girls, three girls), (worst quality:2), (low quality:2),(monochrome:1.1), (normal quality:2), ((grayscale)), (multiple picture:1.3), (multiple photo:1.5), ((wrong feet)),(wrong shoes),(multiple limbs:1.2), (fat ass), lowers, paintings, sketches, skin spots, acnes, skin blemishes, age spot, glans, skin blemishes, bad feet, bad hands, distorted, blurry, missing fingers, multiple feet, bad knees, extra fingers, bad body, glans, nsfw, saggy breast, acnes, age spot, dark spots, fat, fused, giantess, glans, mole, obesity, skin blemishes, skin spots, animal ears, elf-ears, earrings, childish, morbid, blurry, paintings, sketch, text, logo, easy negative,, worst face, error,, horror, bad anatomy, multiple arms, deformed fingers, extra legs, third feet, mutated hands, ugly, feet, toes

 

Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 7, Seed: 1948350849, Face restoration: CodeFormer, Size: 512x768, Model hash: e15f78f212, Model: GenesisLifeMix_PrunedFp16VAE, Denoising strength: 0.7, Clip skip: 2, Hires upscale: 2, Hires upscaler: Latent, Version: v1.3.2

Used embeddings: ng_deepnegative_v1_75t [1a3e]
Used embeddings: ulzzang-6500-v1.1 [873c]

2. 모델 A (Seed 번호 : 1948350849) 

Upscale by 1.0

Upscale by 1.2
Upscale by 1.5
Upscale by 2.0

3. 모델 B (Seed 번호 : 2195516486)

Upscale by 1.0
Upscale by 1.2
Upscale by 1.5
Upscale by 2.0

4. 모델 C (Seed 번호 : 99463579)

Upscale by 1.0
Upscale by 1.2
Upscale by 1.5
Upscale by 2.0

5. 모델 D (Seed 번호 : 1516952835)

Upscale by 1.0
Upscale by 1.2
Upscale by 1.5
Upscale by 2.0

6. 모델 E (Seed 번호 : 32976302)

Upscale by 1.0
Upscale by 1.2
Upscale by 1.5
Upscale by 2.0

7. 모델 F (Seed 번호 : 398685448)

Upscale by 1.0
Upscale by 1.2
Upscale by 1.5
Upscale by 2.0

 

 

반응형