Decode intermediate steps

[oskarjor]

What is the best way to decode intermediate steps at their native resolution, i.e. when generating an image decode resolutions such as 64x64 pixels and 128x128 pixels? I've realised I cannot use VQVAE.fhat_to_image as that only yields images at max scale, and using VQVAE.embed_to_image yielded very poor results.

results = self.vae_proxy[0].embed_to_img(ms_h_BChw, all_to_max_scale=False, last_one=False) 
results = [result.add_(1).mul_(0.5) for result in results]

Is there a better way to get good intermediate results? Alternatively, would you be able to share checkpoints for smaller resolution models if available?