About COCO reconstruction MSE using LDM AE
zhihongp opened this issue · comments
As reported in Table 1, the MSE is 0.015. But I get 0.022 using the same settings as yours based on understand: using stable diffusion v1.4, 5000 COCO val images and image normalized to (-1, 1). Is there anything I am missing here?
Hello, thanks for your interest! If you post your code I can check it out after the ICML deadline next week and post our reconstruction code/compare? You can also experiment with our codes recon error by use EDICT_editing with init_image_strength=0 to just encode/decode with the VAE iirc.
Believe I have found the reason. Instead of using central crop and resize to 512x512, I didn't crop and resized w,h so that %64=0
And you have to use NN for Image.resize(), otherwise it's MSE=0.018 instead of 0.015