512 x 512로 뽑고 hires 4배 돌려봤음.
사진 흐려서 복붙하면
(masterpiece:1.05), (finely detailed:1.05), (intricate details:1.05), (Impeccable attention to detail:1.05), (masterful execution:1.05), (stunning beauty:1.05), (Exceptional quality:1.05), (flawless craftsmanship:1.05), (expressive artistry:1.05), <lora:디테일 조정:1.5>,
Negative prompt: (worst quality, low quality:1.4), watermark, artist name, (dynamic pose:1.5),
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 737750622, Size: 512x512, Model hash: d17ace4d35, Denoising strength: 0.5, Clip skip: 2, Hires upscale: 4, Hires upscaler: R-ESRGAN 4x+ Anime6B, Lora hashes: "디테일 조정: 7c6bad76eb54", Version: v1.3.1
processing | 262.9/13.5s
소요 시간 : 2m 38.54sTorch active/reserved: 16740/20712 MiB, Sys VRAM: 10240/10240 MiB (100.0%)
512x512 잘 돌아가는거 같아서 디디테일러 적용해보면
(masterpiece:1.05), (finely detailed:1.05), (intricate details:1.05), (Impeccable attention to detail:1.05), (masterful execution:1.05), (stunning beauty:1.05), (Exceptional quality:1.05), (flawless craftsmanship:1.05), (expressive artistry:1.05), <lora:디테일 조정:1.5>,
Negative prompt: (worst quality, low quality:1.4), watermark, artist name, (dynamic pose:1.5),
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 1184018342, Size: 512x512, Model hash: d17ace4d35, Denoising strength: 0.4, Clip skip: 2, Lora hashes: "디테일 조정: 7c6bad76eb54", Version: v1.3.1
소요 시간 : 2m 54.32s
Torch active/reserved: 16636/20940 MiB, Sys VRAM: 10240/10240 MiB (100.0%)
안정적으로 버텨줌.
768x512 로 4배 돌리면
OutOfMemoryError: CUDA out of memory. Tried to allocate 6.00 GiB (GPU 0; 10.00 GiB total capacity; 14.49 GiB already allocated; 0 bytes free; 17.44 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
소요 시간 : 5m 22.47sTorch active/reserved: 16414/21402 MiB, Sys VRAM: 10240/10240 MiB (100.0%)
가 나옴.
그래서 3.5배로 내려보고 돌려봤음.
(masterpiece:1.05), (finely detailed:1.05), (intricate details:1.05), (Impeccable attention to detail:1.05), (masterful execution:1.05), (stunning beauty:1.05), (Exceptional quality:1.05), (flawless craftsmanship:1.05), (expressive artistry:1.05), <lora:디테일 조정:1.5>,
Negative prompt: (worst quality, low quality:1.4), watermark, artist name, (dynamic pose:1.5),
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 1928241950, Size: 768x512, Model hash: d17ace4d35, Denoising strength: 0.4, Clip skip: 2, Lora hashes: "디테일 조정: 7c6bad76eb54", Version: v1.3.1
소요 시간 : 4m 11.22s
Torch active/reserved: 18851/21678 MiB, Sys VRAM: 10240/10240 MiB (100.0%)
원래 내 설정인 오토좌 0.64 + 엔비디아 531 에서 512x512는 hires 2.5가 한계고 768x512는 hires 2.0이 한계였는데 512x가 4배, 768이 3.5배까지 먹히는거 보면 시스템 메모리 땡겨오는게 확실하게 효과가 있음.
버그는 아직까지 모르겠음. 컴퓨터 여러대라 그림뽑을땐 다른거 안건들였거든.
