class wise AdaIN

[kimsekeun]

Thank you for great work,

I have two questions.

1. do we need class masks for model input?
2. If we do not have class mask, then can we use this model ? e.g. disable class wise adain then it works?

• i think this is " Adain Diffusion" described in paper, how can i use this model to compare with proposed one?

3. In figure 1, DDPM inversion process, do we need to pretrained DDPM for input modality?

[echigot]

Hello,

If you don't have class masks, the filtering module of CACTIF will still work with standard AdaIN instead of class-wise AdaIN. However, the current implementation asks for labels so you would have to disable manually some parts of the code. A lazy workaround would be to copy/paste any image to your style/labels folder and rename it with a valid label name (Cityscapes or ACDC format). Ex: copy/paste GOPR0122_frame_000234_rgb_anon.png from the images folder to the labels folder, and rename it GOPR0122_frame_000234_gt_labelColor.png. You could then run the code with the --adain_class False option.

For the DDPM inversion process, you don't need to change or download anything more than the standard Stable Diffusion v1.5.

I hope this clarifies your interrogations.