side 1024, tile 512, overlap 64: starts [0, 448, 896], lengths [512, 512, 128], 3 tiles
side 1536, tile 512, overlap 64: starts [0, 448, 896, 1344], lengths [512, 512, 512, 192], 4 tiles
side 960, tile 512, overlap 64: starts [0, 448], lengths [512, 512], 2 tiles
ceil((D - O) / (S - O)) agrees with the number of tile starts for every D from 100 to 700, tiles of 64, 128 and 256, overlaps 0, 8 and 32

input side   tiles per side   tiles   pixels fed to the model / pixels in the image
512          1                1       1.000
768          2                4       1.085
1024         3                9       1.129
2048         5                25      1.129

VAEDecodeTiled on a 1024 x 1024 image: latent 128x128, tiles ((32, 128), (128, 32), (64, 64)), decoder calls [5, 5, 9], 19 in all, against 9 for the (64, 64) tiling alone
latent positions decoded: 61,696 for 16,384 in the latent, 3.77 times the work of one untiled pass

latent 64x64, tile 32, receptive field 3 latent pixels. Largest error of the tiled output against the untiled output, as a multiple of the output's standard deviation (0.308)
overlap   tiles   pixels fed in / pixels   plain average   feathered (ComfyUI's mask)   three tilings averaged
0         4       1.000                  2.76            2.76                         1.84
2         9       1.129                  1.90            1.61                         0.94
4         9       1.266                  1.71            0.98                         0.50
8         9       1.562                  1.51            0.39                         0.29
