Plane 0
3D multi-plane: binary amplitude vs binary phase-only
DragonBunny RGB-D, 9 focal planes, one shared 8-frame hologram set, 300 iterations. Each hardware mode independently selects LR from the same five-point 100-iteration screen. Targets are loaded as linear optical amplitude; display previews square to intensity and apply sRGB only for viewing.
0.1selected LR, both modes
28.06 dBFLCOS mean PSNR
25.25 dBphase-only mean PSNR
1.63×phase/L2 amplitude error
Source RGB-D
Independent LR screen
Quantitative comparison
| Mode | LR | mean amp MSE | mean PSNR | mean SSIM | focus Δ RMSE | plane spread ratio |
|---|---|---|---|---|---|---|
| FLCOS binary amplitude | 0.1 | 0.00355 | 28.06 | 0.3873 | 0.05076 | 2.247 |
| Binary phase-only (0/π) | 0.1 | 0.00578 | 25.25 | 0.3143 | 0.04786 | 1.993 |
Focus Δ RMSE evaluates differences relative to the middle focal plane. Plane-spread ratio is reconstructed focal variation divided by target variation: 1 is the correct variation magnitude, while a low value indicates collapsed/blurred depth behavior.
First finding
True 0/π phase-only modulation is substantially more constrained than binary amplitude in this forward model. Even after independent LR tuning, it retains more speckle and cross-plane residual. This is a physically meaningful hard setting—not an artificial downsampling—and gives the new 3D loss room to improve focus allocation and suppress structured noise.
Plane 1
Plane 2
Plane 3
Plane 4
Plane 5
Plane 6
Plane 7
Plane 8
