체크포인트 모델(Check Point)별 실사이미지 생성

스테이블 디퓨전에서 동일한 Prompt를 적용한 상태에서
체크포인트 모델들을 변경하면 어떤 이미지가 생성되는지 테스트한 것입니다.
조건은 동일한 Prompt를 적용하고 한 번은 체크포인트 모델만 변경하면서 이미지를 생성하고,
한 번은 한 개의 조합 로라(Lora) 모델을 적용하고 체크포인트 모델을 변경하면서 이미지를 생성하였습니다.

차이는 있겠지만,
먼저 체크포인트 모델을 어떤 것을 적용하느냐에 따라 실사이미지의 차이를 확인할 수 있습니다.
두 번째는 어떤 로라(Lora) 모델을 적용하느냐에 따라 더 많은 차이를 확인할 수 있습니다.
생성하는 이미지가 무엇이냐에 따라 다르겠지만,
좋은 체크포인트 모델을 선택하고 체크포인트에 어울리는 로라 모델을 선택한 후 로라 모델의 적용 비율을 잘 조절하면 좀 더 실사이미지에 가까운 결과물을 얻을 수 있을 것 같습니다.

 

제목 : 체크포인트 모델(Check Point)별 실사이미지 생성, #실사이미지 - YouTube

 

===================================================================

 

[ Prompt ]
best quality, best shadow, masterpiece, realistic, photo-realistic, realism, extremely detailed CG, Amazing, finely detail, Super Resolution, super detail, ultra-detailed, ultra high res, finely detail, beautiful detailed sky, RAW photo, detailed cafe, HDR, realistic skin texture, real human skin, highly detailed skin, beautiful white shiny humid skin, shiny skin, perfect fingers, depth of field, 4k, 8k, 16k
(1 girls:1.5), (20 age:1.2), korean pretty women, korean beautiful women, smile, large breasts, korea style dress, (dress with small pattern:1.3), colored dress, red dress|blue dress|orange dress|green dress|pink dress, high heel, standing, (white background:1.5),

[ Negative prompt ]
paintings, sketches, (worst quality:2), (low quality:2), lowres, ((monochrome)), ((grayscale)), skin spots, acnes, skin blemishes, age spot, (ugly:1.331), (duplicate:1.331), (mutilated:1.21), mutated hands, (poorly drawn hands:1.331), (bad anatomy:1.21), (bad proportions:1.331), extra limbs, (missing arms:1.331), (extra legs:1.331), (fused fingers:1.61051), (too many fingers:1.61051), (unclear eyes:1.331), missing fingers, extra digit, (futa:1.1), bad body, NG_DeepNegative_V1_75T, negative_handbag

 

[ 설  정 ]
Steps : 25
Sampler : DPM++ SDE Karras
CFG scale : 7
Seed : 1695371280
Size : 512x768
DDetailer model a : bbox\mmdet_anime-face_yolov3.pth [51e1af4a]
DDetailer model b : segm\mmdet_dd-person_mask2former.pth [3848e764]
Hires upscale : 2
Hires steps : 25
Hires upscaler : Latent (bicubic antialiased)

로라(Lora) : <lora:koreaface15:0.5> <lora:AAV-tsukasa:0.2> <lora:fish1v1:0.3>

[Model Link]
01. beautifulRealistic_v60 : https://civitai.com/models/25494?modelVersionId=113479 
02. beautifulRealistic_brav5 : https://civitai.com/models/25494?modelVersionId=63786 
03. chilloutmix_NiPrunedFp32Fix : https://civitai.com/models/6424/chilloutmix
04. consistentFactor_euclidV51 : https://civitai.com/models/9114?modelVersionId=89824 
05. dream2reality_v10 : https://civitai.com/models/105573/dream2reality
06. goodAsianGirlFace_goodAsianGirlFaceV12 : https://civitai.com/models/5726/good-asian-girl-face
07. gracefulmix_v10 : https://civitai.com/models/46959/gracefulmix
08. kisaragiMix_v22 : https://civitai.com/models/45757/kisaragimix
09. majicmixRealistic_v6 : https://civitai.com/models/43331/majicmix-realistic
10. Moonfilm_filmGrain10 : https://civitai.com/models/43977?modelVersionId=92475 
11. Moonfilm_reality30 : https://civitai.com/models/43977?modelVersionId=73071 
12. nobleMixFix_v30 : https://civitai.com/models/65480/noble-mix-fix
13. simplemix_V2 : https://civitai.com/models/62921?modelVersionId=76122 

[Lora Link] 
koreaface15 : https://civitai.com/models/77710/normal-korean-girl-face-chilloutmix-base-lora
AAV-tsukasa : https://civitai.com/models/35216/wrav-tsukasa
fish1v1 : https://civitai.com/models/12132/fish-lora

 

01; beautifulRealistic_v60

02. beautifulRealistic_brav5

03. chilloutmix_NiPrunedFp32Fix

04. consistentFactor_euclidV51

05. dream2reality_v10

06. goodAsianGirlFace_goodAsianGirlFaceV12

07. gracefulmix_v10

08. kisaragiMix_v22

09. majicmixRealistic_v6

10. Moonfilm_filmGrain10

11. Moonfilm_reality30

12. nobleMixFix_v30

13. simplemix_V2