12 channels

full encoder 512, 16L

lightweight encoder

retrain with smaller batch size and more extreme lr decay

16 channels 256 encoder

256 encoder 8 layer decoder

allow 1, 3, or 4 channels for entropy coding

webp entropy coding

README

README

save as multi-channel image

extended training

sd3 autoencoder without tft

(tfabs) phi = 0.01, latent_dim=16, J=3, 5 bits

(tf) phi = 1.0, latent_dim=16, J=3, 5 bits

metadata

Update README.md

entropy coding

phi = 0, latent_dim=16, J=3, 5 bits

Update README.md

example usage

example usage

(tfnorm) phi = 0.01, latent_dim=16, J=3, 5 bits

(tfnorm) phi = 0.01, latent_dim=48, J=4, 6 bits

(tfnorm) phi = 0, latent_dim=48, J=4, 6 bits

phi = 0.1, latent_dim=12, J=3, 5 bits

(d)phi = 0.01, latent_dim=9, J=3, 4 bits

70 bits (16c 5b)

36 bits (12c 3b)

actually 36 bits (6c 6b)

36 bits (6c 6b)

16x rr

LSDIR 540

no rr, 768

bior 4.4

no rr 1536

deeper (6 instead of 4)

8 bits

32x rr 48x dr (baseline)

true wavelet with hypercodec, modified quantization, 8x rr

symmetric fb, ternary quantization, 10d (16 bits)

true wavelet with hypercodec, 8x rr, 12x dr

symmetric, 48x dr

symmetric, 12x dr

true wavelet, 12x dr, with hypercodec 15 epochs

true wavelet transform, wavelet loss, 12.8x dr, 12 channels per scale, 8 bits, no hypercodec, 15 epochs

true wavelet transform, 12.8x dr, 12 channels per scale, 8 bits, no hypercodec, 15 epochs

initial commit