KL(N(1.2,0.6^2) || N(0,1))  closed form 0.9108   Monte Carlo 0.9102

beta    reconstruction MSE per image   KL per image   latent std
0.0     0.0002                         18.562         2.909
0.001   0.0019                         9.058          1.459
0.05    0.0058                         4.102          1.063
1.0     0.0497                         1.704          0.690

latent std per channel with a weak KL: 1.524 1.381   scale factor 1/std: 0.6853   std after scaling: 1.000

one row of pixels
image A                  0.10 0.77 0.77 0.10 0.00 0.00 0.00 0.00
image B                  0.00 0.00 0.00 0.00 0.10 0.77 0.77 0.10
pixel midpoint           0.05 0.39 0.39 0.05 0.05 0.39 0.39 0.05
true bar at the middle   0.00 0.00 0.10 0.77 0.77 0.10 0.00 0.00
latent midpoint, beta 0.001 0.05 0.34 0.33 0.14 0.11 0.26 0.45 0.11
latent midpoint, beta 0.05  0.06 0.31 0.20 0.00 0.13 0.61 0.41 0.03
latent midpoint, beta 1.0   -0.00 -0.00 0.12 0.72 0.77 0.15 0.01 0.01
distance from a real bar at the middle:  pixel midpoint 12.115   beta 0.001: 9.762   beta 0.05: 12.322   beta 1.0: 0.047

format          factor  channels  values in 1024x1024   compression
SD 1.x, SDXL    8       4         65,536                48x
SD3, FLUX.1     8       16        262,144               12x
