"FouriScale: A Frequency Perspective on Training-Free High-Resolution Image Synthesis" ECCV, 2024 Mar 19 paper code pdf note Authors: Linjiang Huang, Rongyao Fang, Aiping Zhang, Guanglu Song, Si Liu, Yu Liu, Hongsheng Li
-
Task
-
Problems
- addressing persistent challenges, such as repetitive patterns and structural distortions,
-
🏷️ Label:
-
we introduce an innovative, training-free approach FouriScale from the perspective of frequency domain analysis.
-
We replace the original convolutional layers in pre-trained diffusion models by incorporating a dilation technique along with a low-pass operation, intending to achieve structural consistency and scale consistency across resolutions, respectively
- ScaleCrafter 直接生成的细节不行
把 UNet 的 Conv Kernel 变成 dilated，去和 Low-frequency 特征相乘
高频细节更多
CFG 使用场景
ablation study 看那个模块有效，总结一下
More importantly, as shown in Fig. 3, we find that our method can generate much better details than the cascaded pipeline. Due to a lack of prior knowledge in generation, super-resolution method can only utilize existing knowledge within the single image for upscaling the image, resulting in over-smooth appearance.
指标没有 SD+SR 高，但看图像细节会清晰超级多，但和原始的一致性不好。后续工作 FAM Diffusion 也验证了本文的一致性较烂
learn what