본문 바로가기

딥러닝/생성AI

(3)
sd, sdxl 아키텍처 https://ostin.tistory.com/231 conv_in down_blocks:    (CrossAttnDownBlock2D:        ResnetBlock2D        Transformer2DModel        ResnetBlock2D        Transformer2DModel        Downsample2D    ) x 3    DownBlock2D:        (ResnetBlock2D) x 2 mid_blocks:        ResnetBlock2D        Transformer2DModel        ResnetBlock2D up_blocks:    UpBlock2D:       (ResnetBlock2D) x 3        Upsample2D    (Cr..
[diffusers] controlnet 코드 분석 코드 매우 심플한데, 효과적인 생성방법. down block, mid block 부분 controlnet에서 떼옴 .  down_block_res_samples, mid_block_res_sample = controlnet(                    noisy_latents,                    timesteps,                    encoder_hidden_states=encoder_hidden_states,                    controlnet_cond=controlnet_image,                    return_dict=False,                )unet 의 down block부분과 mid block부분을 업데이..
[prompt to prompt] 설치하기 & 에러들 1) attentionReplace 2) Refinement : 단어가 추가되어 refinement 필요 3) reweight => 일단 궁금한것 refinement!! 단어길이가 다른데, 어떻게 처리했음? Attention Control Options 더보기 cross_replace_steps: specifies the fraction of steps to edit the cross attention maps. Can also be set to a dictionary [str:float] which specifies fractions for different words in the prompt. self_replace_steps: specifies the fraction of steps to repla..