Hello, thanks for the great work, I learned a lot from the paper!
I have a question about the tiled_encode implementation in the VAE encoder.
In tile_parallel.py, each tile is encoded separately using self.encode_fn with its default parameters. From the defaults here, sample_posterior=True, which means the encoding step produces the distribution parameters and then samples from it.
Later, in tile_parallel.py, these sampled latents are blended across tiles. This is a sample-then-blend approach, whereas many other VAE tiling implementations follow a blend-then-sample pattern (blending means/variances first, then sampling once from the blended distribution).
Is there a specific reason you chose sample-then-blend instead of blend-then-sample? I’m curious if it was for performance, simplicity, or a particular modeling choice.
Hello, thanks for the great work, I learned a lot from the paper!
I have a question about the
tiled_encodeimplementation in the VAE encoder.In tile_parallel.py, each tile is encoded separately using
self.encode_fnwith its default parameters. From the defaults here,sample_posterior=True, which means the encoding step produces the distribution parameters and then samples from it.Later, in tile_parallel.py, these sampled latents are blended across tiles. This is a sample-then-blend approach, whereas many other VAE tiling implementations follow a blend-then-sample pattern (blending means/variances first, then sampling once from the blended distribution).
Is there a specific reason you chose sample-then-blend instead of blend-then-sample? I’m curious if it was for performance, simplicity, or a particular modeling choice.