0 = 8x8 16x16 ;Caution:
1 = 8x8 32x32 ;In 224-lines mode, OBJs with 64-pixel height
2 = 8x8 64x64 ;may wrap from lower to upper screen border.
3 = 16x16 32x32 ;In 239-lines mode, the same problem applies
4 = 16x16 64x64 ;also for OBJs with 32-pixel height.
5 = 32x32 64x64
6 = 16x32 32x64 (undocumented)
7 = 16x32 32x32 (undocumented)
You can only have two sprite sizes on screen at the same time on the SNES (for that frame). That can be pretty limiting. If you want to optimize for sprite scanline flicker (overflow) and use 8x8 for one size, you're forced to use 16x16 or 32x32 or 64x64 for the other size. That last two pairs are rather useless for the most part IMO. So 8x8 paired with 16x16. A little limiting and it'll eat through the 128 entry OAM (sprite table) quicker.
If you optimized for 8x8/16x16, then you hit the sweet spot for sprite cells per scanline limit. It's 34 8x8 cells per scanline or 272 pixels (byuu corrected me a couple of years back). I think most SNES games used 16x16/32x32 sprite mode. Which is pretty close to PCE for scanline pixel to cell limit. But the SNES has the advantage for another BG layer (and a 3rd low color one in some modes) to make larger enemies/bosses without cutting into that sprite scanline limit.
There's a hardware trick on the PCE that you can do to make 16x8 sprite cells ( making the following sizes 16x8,32x8,16x16,32x16,16x32,32x32) instead of the usual 16x16 (16x16,32x16,16x32,32x32,16x64,32x64). Not as nice as 8x8, but still makes it more optimal in certain situations since it actually effects the scanline limit on a vertical plane/path.