https://github.com/lllyasviel/ControlNet-v1-1-nightly#controlnet-11-tile
tile model
control_v11f1e_sd15_tile, from https://huggingface.co/lllyasviel/ControlNet-v1-1/tree/main
tile model 能够生成新的图像细节，并忽略原图里的已有图像细节，因此，该模型可以用于移除不好的细节，添加改善后的细节. 例如，移除由于图像缩放所引起的模糊.
参考：超分
但，并不是超分模型.
其主要有两个功能：
- 忽略图片的细节，生成新的细节
- 若 local tile semantics 和 prompts 不匹配，则忽略 global prompts，并采用 local context 来引导 diffusion 生成.
如果原图像本身已经有很好的细节，仍然可以采用 tile model 替换图像细节. Stable Diffusion's I2I 可以实现类似作用，但 tile model 更易于保持图像的整体结构，只改变细节，即使 denosing strength =1.0.
ControlNet tile model，对于给定 tile，其首先识别 tile 里所包含的东西，并增强其识别到的语义的影响；且，当 内容与 prompts 不匹配时，降低 global prompts 的影响.
sd-webui-controlnet tile model
使用方式：
- it can do 2x, 4x, or 8x super resolution
- it can add, or change, or re-generate image details in an image
- it can fix, refine, and improve bad image details obtained by any other super resolution methods like bad details or blurring from RealESRGAN
- it can guide SD to diffuse in tiles, "one beautiful girl" will not generate 16 girls if you use 16 tiles and denoising strength 1.0.
- it can finish unfinished artwork drafts if those drafts are drawn by color blocks
Preprocessor:
- The preprocessor can be none or tile_resample (tile_resample is just a resizing to make your image smaller if you want more variations)
Hint:
- do not forget to try different control modes - it can produce interesting effects
- this model supports resolution 768 or 1024 well. You can try it if you find SD's 512 is too small and blur.