Dataset Viewer
Auto-converted to Parquet Duplicate
model
stringclasses
5 values
model_label
stringclasses
5 values
window
int64
1
4
resolution
stringclasses
5 values
backend
stringclasses
5 values
precision_name
stringclasses
2 values
compile
bool
2 classes
median_ms
float64
6.69
1.23k
fps
float64
0.81
149
min_ms
float64
6.49
1.05k
⌀
p90_ms
float64
11
769
⌀
max_ms
float64
7.22
1.1k
⌀
spread_pct
float64
0.6
70
⌀
timed_n
float64
10
20
⌀
load_s
float64
0
2.92
⌀
compile_s
float64
0.02
5.89
⌀
peak_mb
float64
359
13.8k
⌀
apple_m4_max_ms
float64
10
1.23k
⌀
source
stringclasses
3 values
status
stringclasses
2 values
fov
FOV
1
896x512
coreml
null
null
27.736
36.054
27.535
null
30.773
11.67
20
0.2453
1.6069
10,267.5
null
phase1
ok
b
B
1
1138x640
coreml
null
null
55.919
17.883
55.667
null
58.844
5.68
20
0.7273
0.2295
11,309.9
85
phase1
ok
b
B
1
1280x720
coreml
null
null
73.785
13.553
73.268
null
87.275
18.98
20
0.7291
0.3777
11,694.4
115
phase1
ok
b
B
1
640x360
coreml
null
null
16.29
61.387
15.999
null
18.587
15.88
20
0.7809
0.1718
10,413.4
26
phase1
ok
b
B
1
854x480
coreml
null
null
28.542
35.036
27.965
null
43.22
53.45
20
0.7444
0.1891
10,651.1
44
phase1
ok
b
B
1
896x512
coreml
null
null
31.521
31.725
31.278
null
47.286
50.79
20
0.7254
0.2301
10,757.2
48
phase1
ok
h
H+
1
1138x640
coreml
null
null
344.994
2.899
331.736
null
370.617
11.27
20
2.4895
0.5784
13,244.3
612
phase1
ok
h
H+
1
1280x720
coreml
null
null
494.775
2.021
476.896
null
519.99
8.71
20
2.4598
0.7049
13,823
826
phase1
ok
h
H+
1
640x360
coreml
null
null
87.494
11.429
86.283
null
89.904
4.14
20
2.4808
0.3499
12,025.6
170
phase1
ok
h
H+
1
854x480
coreml
null
null
163.68
6.109
156.066
null
181.105
15.3
20
2.4719
0.433
12,345.3
303
phase1
ok
h
H+
1
896x512
coreml
null
null
192.944
5.183
180.74
null
203.4
11.74
20
2.491
0.4564
12,417.7
339
phase1
ok
l
L
1
1138x640
coreml
null
null
169.458
5.901
164.019
null
183.37
11.42
20
1.5764
0.3607
11,993.5
272
phase1
ok
l
L
1
1280x720
coreml
null
null
242.02
4.132
232.077
null
257.623
10.56
20
1.5773
0.398
12,457.2
385
phase1
ok
l
L
1
640x360
coreml
null
null
45.827
21.821
45.674
null
47.989
5.05
20
1.5693
0.2159
10,982.7
80
phase1
ok
l
L
1
854x480
coreml
null
null
82.776
12.081
81.753
null
97.439
18.95
20
1.5847
0.2556
11,315.5
137
phase1
ok
l
L
1
896x512
coreml
null
null
94.098
10.627
91.774
null
108.49
17.76
20
1.5675
0.307
11,441.6
152
phase1
ok
s
S
1
1138x640
coreml
null
null
22.932
43.607
22.78
null
36.857
61.39
20
0.606
0.1769
10,703.6
34
phase1
ok
s
S
1
1280x720
coreml
null
null
31.483
31.763
31.243
null
45.029
43.79
20
0.5998
0.1863
10,924.5
45
phase1
ok
s
S
1
640x360
coreml
null
null
6.691
149.444
6.494
null
7.219
10.83
20
0.588
0.334
10,209.5
10
phase1
ok
s
S
1
854x480
coreml
null
null
11.879
84.181
11.747
null
14.21
20.73
20
0.6334
0.1663
10,348.4
17
phase1
ok
s
S
1
896x512
coreml
null
null
13.197
75.774
12.973
null
15.727
20.86
20
0.6143
0.3948
10,378.8
19
phase1
ok
b
B
4
1138x640
coreml
null
null
136.729
7.314
133.566
null
142.508
6.54
20
0.836
0.323
5,215.1
137
phase1
ok
b
B
4
1280x720
coreml
null
null
206.705
4.838
204.05
null
212.849
4.26
20
0.8309
0.5097
8,633.3
211
phase1
ok
b
B
4
640x360
coreml
null
null
25.859
38.672
25.111
null
27.835
10.53
20
0.8451
1.9383
3,661
32
phase1
ok
b
B
4
854x480
coreml
null
null
53.687
18.627
52.42
null
55.423
5.59
20
0.8646
1.8648
2,271.1
60
phase1
ok
b
B
4
896x512
coreml
null
null
60.92
16.415
59.829
null
63.958
6.78
20
0.8358
0.9726
2,974.6
68
phase1
ok
h
H+
4
1138x640
coreml
null
null
709.185
1.41
687.909
null
726.949
5.5
20
2.9169
0.9582
11,753.4
832
phase1
ok
h
H+
4
1280x720
coreml
null
null
1,081.809
0.924
1,054.862
null
1,101.747
4.33
20
2.8104
1.3784
8,408.6
1,233
phase1
ok
h
H+
4
640x360
coreml
null
null
140.259
7.13
133.776
null
145.935
8.67
20
2.9076
0.849
2,930.5
197
phase1
ok
h
H+
4
854x480
coreml
null
null
290.804
3.439
276.44
null
298.543
7.6
20
2.881
0.561
5,032.3
375
phase1
ok
h
H+
4
896x512
coreml
null
null
334.388
2.991
319.522
null
342.944
7
20
2.8484
0.6213
5,679.4
424
phase1
ok
l
L
4
1138x640
coreml
null
null
390.187
2.563
379.184
null
406.543
7.01
20
2.2824
0.6054
7,592.5
425
phase1
ok
l
L
4
1280x720
coreml
null
null
606.695
1.648
587.376
null
635.081
7.86
20
2.3234
1.3494
12,133.1
648
phase1
ok
l
L
4
640x360
coreml
null
null
70.11
14.263
69.306
null
72.817
5.01
20
2.3704
1.4114
1,513.9
96
phase1
ok
l
L
4
854x480
coreml
null
null
146.686
6.817
145.119
null
152.017
4.7
20
2.3087
1.4369
2,964.3
188
phase1
ok
l
L
4
896x512
coreml
null
null
172.881
5.784
168.182
null
182.958
8.55
20
2.3397
0.8247
3,766.4
212
phase1
ok
s
S
4
1138x640
coreml
null
null
62.164
16.086
61.234
null
64.675
5.53
20
0.7312
0.7026
3,229.7
61
phase1
ok
s
S
4
1280x720
coreml
null
null
96.854
10.325
95.879
null
102.495
6.83
20
0.7131
1.469
5,547.6
88
phase1
ok
s
S
4
640x360
coreml
null
null
11.209
89.216
11.014
null
13.63
23.34
20
0.7755
5.8914
358.9
13
phase1
ok
s
S
4
854x480
coreml
null
null
25.661
38.97
24.863
null
27.814
11.5
20
0.7759
2.2199
1,045
26
phase1
ok
s
S
4
896x512
coreml
null
null
29.216
34.228
28.371
null
31.104
9.35
20
0.7089
0.9319
1,538.1
30
phase1
ok
b
B
1
1138x640
pytorch_mps_bf16
null
null
86.301
11.587
84.532
null
101.482
19.64
20
0.3115
0.156
2,470
null
phase1
ok
b
B
1
1280x720
pytorch_mps_bf16
null
null
112.668
8.876
110.611
null
116.173
4.94
20
0.3115
0.1563
2,470.6
null
phase1
ok
b
B
1
640x360
pytorch_mps_bf16
null
null
32.798
30.49
31.935
null
38.223
19.17
20
0.3115
0.4225
1,454.3
null
phase1
ok
b
B
1
854x480
pytorch_mps_bf16
null
null
48.24
20.729
47.404
null
64.025
34.46
20
0.3115
0.6617
1,454.8
null
phase1
ok
b
B
1
896x512
pytorch_mps_bf16
null
null
53.184
18.803
52.588
null
59.341
12.7
20
0.3115
0.1092
1,445.4
null
phase1
ok
h
H+
1
1138x640
pytorch_mps_bf16
null
null
474.986
2.105
471.618
null
486.787
3.19
10
1.6035
0.4942
7,271.9
null
phase1
ok
h
H+
1
1280x720
pytorch_mps_bf16
null
null
621.694
1.609
611.974
null
625.943
2.25
10
1.6035
0.6504
8,296.5
null
phase1
ok
h
H+
1
640x360
pytorch_mps_bf16
null
null
162.02
6.172
160.569
null
173.588
8.04
20
1.6035
0.4269
6,238.2
null
phase1
ok
h
H+
1
854x480
pytorch_mps_bf16
null
null
272.724
3.667
270.8
null
273.521
1
10
1.6035
0.497
6,246.8
null
phase1
ok
h
H+
1
896x512
pytorch_mps_bf16
null
null
291.799
3.427
290.222
null
304.516
4.9
10
1.6035
0.3255
7,271.4
null
phase1
ok
l
L
1
1138x640
pytorch_mps_bf16
null
null
268.879
3.719
265.08
null
270.695
2.09
10
0.6948
0.2957
4,195.9
null
phase1
ok
l
L
1
1280x720
pytorch_mps_bf16
null
null
353.618
2.828
349.937
null
364.006
3.98
10
0.6948
0.3695
4,196.5
null
phase1
ok
l
L
1
640x360
pytorch_mps_bf16
null
null
89.004
11.235
87.097
null
95.937
9.93
20
0.6948
0.9662
3,174.2
null
phase1
ok
l
L
1
854x480
pytorch_mps_bf16
null
null
139.654
7.161
137.953
null
141.843
2.79
20
0.6948
1.0457
3,170.8
null
phase1
ok
l
L
1
896x512
pytorch_mps_bf16
null
null
157.171
6.362
152.071
null
159.689
4.85
20
0.6948
0.5299
3,171.4
null
phase1
ok
s
S
1
1138x640
pytorch_mps_bf16
null
null
41.253
24.241
40.36
null
47.453
17.19
20
0.4147
0.5204
1,420
null
phase1
ok
s
S
1
1280x720
pytorch_mps_bf16
null
null
52.621
19.004
51.3
null
62.088
20.5
20
0.4147
0.8323
1,420.6
null
phase1
ok
s
S
1
640x360
pytorch_mps_bf16
null
null
17.76
56.307
17.159
null
22.342
29.19
20
0.4147
3.2204
1,418.3
null
phase1
ok
s
S
1
854x480
pytorch_mps_bf16
null
null
24.542
40.746
23.639
null
27.956
17.59
20
0.4147
1.6629
1,418.8
null
phase1
ok
s
S
1
896x512
pytorch_mps_bf16
null
null
27.978
35.743
26.566
null
31.82
18.78
20
0.4147
0.2113
1,419.4
null
phase1
ok
b
B
4
1138x640
pytorch_mps_bf16
null
null
99.988
10.001
98.03
null
109.204
11.18
20
0.3115
0.089
2,470.6
null
phase1
ok
b
B
4
1280x720
pytorch_mps_bf16
null
null
135.444
7.383
133.372
null
137.574
3.1
20
0.3115
0.115
3,494.6
null
phase1
ok
b
B
4
640x360
pytorch_mps_bf16
null
null
34.337
29.123
33.305
null
49.727
47.83
20
0.3115
0.1511
2,470.6
null
phase1
ok
b
B
4
854x480
pytorch_mps_bf16
null
null
52.755
18.955
51.752
null
62.597
20.56
20
0.3115
0.0507
2,470.6
null
phase1
ok
b
B
4
896x512
pytorch_mps_bf16
null
null
59.145
16.908
57.432
null
73.787
27.65
20
0.3115
0.0558
2,470.6
null
phase1
ok
h
H+
4
1138x640
pytorch_mps_bf16
null
null
521.365
1.918
517.832
null
533.611
3.03
10
1.6035
0.4656
9,320.5
null
phase1
ok
h
H+
4
1280x720
pytorch_mps_bf16
null
null
706.05
1.416
700.646
null
723.655
3.26
10
1.6035
0.603
9,320.5
null
phase1
ok
h
H+
4
640x360
pytorch_mps_bf16
null
null
175.556
5.696
173.324
null
190.203
9.61
20
1.6035
0.1723
8,296.5
null
phase1
ok
h
H+
4
854x480
pytorch_mps_bf16
null
null
284.657
3.513
283.326
null
286.426
1.09
10
1.6035
0.2692
8,296.5
null
phase1
ok
h
H+
4
896x512
pytorch_mps_bf16
null
null
313.337
3.191
312.329
null
314.31
0.63
10
1.6035
0.294
8,296.5
null
phase1
ok
l
L
4
1138x640
pytorch_mps_bf16
null
null
307.256
3.255
304.296
null
329.159
8.09
10
0.6948
0.255
5,228.5
null
phase1
ok
l
L
4
1280x720
pytorch_mps_bf16
null
null
414.745
2.411
408.388
null
421.516
3.17
10
0.6948
0.3461
6,252.5
null
phase1
ok
l
L
4
640x360
pytorch_mps_bf16
null
null
96.147
10.401
93.57
null
106.73
13.69
20
0.6948
0.0926
4,204.5
null
phase1
ok
l
L
4
854x480
pytorch_mps_bf16
null
null
165.15
6.055
160.168
null
178.689
11.21
20
0.6948
0.1443
4,204.5
null
phase1
ok
l
L
4
896x512
pytorch_mps_bf16
null
null
181.49
5.51
176.557
null
184.952
4.63
20
0.6948
0.1565
5,228.5
null
phase1
ok
s
S
4
1138x640
pytorch_mps_bf16
null
null
46.127
21.679
44.296
null
48.576
9.28
20
0.4147
0.0425
1,428.6
null
phase1
ok
s
S
4
1280x720
pytorch_mps_bf16
null
null
61.186
16.344
59.34
null
63.574
6.92
20
0.4147
0.0524
1,428.6
null
phase1
ok
s
S
4
640x360
pytorch_mps_bf16
null
null
19.163
52.184
18.35
null
26.066
40.27
20
0.4147
0.1335
1,428.6
null
phase1
ok
s
S
4
854x480
pytorch_mps_bf16
null
null
26.199
38.169
25.328
null
35.873
40.25
20
0.4147
0.0325
1,428.6
null
phase1
ok
s
S
4
896x512
pytorch_mps_bf16
null
null
28.802
34.72
28.241
null
31.161
10.14
20
0.4147
0.0282
1,428.6
null
phase1
ok
fov
FOV
1
896x512
mlx_bf16
bf16
true
15.656
63.875
15.473
15.736
15.835
2.32
20
0.0028
0.428
1,034.1
null
phase2
ok
b
B
1
1138x640
mlx_bf16
bf16
true
77.98
12.824
76.116
78.829
79.148
3.89
20
0.0039
0.0792
2,256.9
85
phase2
ok
b
B
1
1280x720
mlx_bf16
bf16
true
102.564
9.75
99.765
103.377
104.241
4.36
20
0.0039
0.1016
2,238.9
115
phase2
ok
b
B
1
640x360
mlx_bf16
bf16
true
25.444
39.302
24.538
25.965
26.004
5.76
20
0.0039
0.0741
1,325.2
26
phase2
ok
b
B
1
854x480
mlx_bf16
bf16
true
41.487
24.104
39.596
41.954
42.096
6.03
20
0.0039
0.0482
1,789.3
44
phase2
ok
b
B
1
896x512
mlx_bf16
bf16
true
46.484
21.513
44.275
47.006
47.139
6.16
20
0.0039
0.0525
1,770.3
48
phase2
ok
h
H+
1
1138x640
mlx_bf16
bf16
true
360.376
2.775
358.778
361.113
361.894
0.86
10
0.0078
0.3739
5,939.2
612
phase2
ok
h
H+
1
1280x720
mlx_bf16
bf16
true
472.693
2.116
470.372
474.888
475.237
1.03
10
0.0078
0.4843
6,594.4
826
phase2
ok
h
H+
1
640x360
mlx_bf16
bf16
true
127
7.874
126.21
127.277
127.553
1.06
20
0.0078
0.285
4,643.1
170
phase2
ok
h
H+
1
854x480
mlx_bf16
bf16
true
200.983
4.976
200.224
201.693
202.107
0.94
20
0.0078
0.2236
5,130.6
303
phase2
ok
h
H+
1
896x512
mlx_bf16
bf16
true
221.083
4.523
220.02
221.681
222.168
0.97
20
0.0078
0.2399
5,239.4
339
phase2
ok
l
L
1
1138x640
mlx_bf16
bf16
true
197.82
5.055
196.671
199.56
199.733
1.55
20
0.0059
0.2085
3,845.4
272
phase2
ok
l
L
1
1280x720
mlx_bf16
bf16
true
259.989
3.846
258.692
261.36
262.126
1.32
10
0.0059
0.2694
4,370.8
385
phase2
ok
l
L
1
640x360
mlx_bf16
bf16
true
69.138
14.464
68.892
69.546
70.623
2.5
20
0.0059
0.0873
2,721.9
80
phase2
ok
l
L
1
854x480
mlx_bf16
bf16
true
108.111
9.25
107.822
108.398
108.534
0.66
20
0.0059
0.124
3,050.3
137
phase2
ok
l
L
1
896x512
mlx_bf16
bf16
true
118.778
8.419
117.933
119.519
119.855
1.62
20
0.0059
0.1343
3,144.5
152
phase2
ok
s
S
1
1138x640
mlx_bf16
bf16
true
30.88
32.384
30.479
31.063
31.103
2.02
20
0.0036
0.0362
1,304.4
34
phase2
ok
s
S
1
1280x720
mlx_bf16
bf16
true
40.748
24.541
40.46
40.95
41.095
1.56
20
0.0036
0.0467
1,622.3
45
phase2
ok
s
S
1
1280x720
mlx_bf16
bf16
false
43.523
22.976
43.336
43.66
43.757
0.97
20
0.0036
0.0484
1,627.4
45
phase2
ok
End of preview. Expand in Data Studio

Apple Depth Pro 2 on a 64GB M5 Pro Mac mini

Headline: I ran Apple Depth Pro 2 on my 64GB M5 Pro Mac mini across three runtimes: Core ML, PyTorch MPS, and an MLX port. At 720p, single-frame H+ took 472.7 ms in MLX BF16, 494.8 ms in Core ML, and 621.7 ms in PyTorch MPS. Apple's published figure for that same cell on a 64GB M4 Max MacBook Pro is 826 ms. That M4 Max number is Apple's, measured on a different machine.

This is an independent local measurement from November 19–20, 2026. I did not upload any original Depth Pro 2 weights in this dataset.

The MLX port (depth_pro_2_mlx 0.1.0, on MLX 0.32.4) was built by @SpaceXSI with a coding agent. The port itself is not in this dataset.

A background inference app was paused at the owner's request during the Phase 1 and Phase 2 speed runs. The later 4-bit quality check left it running.

Apple Machine Learning Research Model is licensed under the Apple Machine Learning Research Model License Agreement. Depth Pro 2 is for research use only. This dataset redistributes measurements, scripts, and charts, not model weights.

Setup

  • Apple M5 Pro Mac mini, 64 GB unified memory, peak memory bandwidth 307 GB/s, macOS 27.0.1 (26A434), AC power.
  • Phase 1 (Core ML + PyTorch MPS): Python 3.12.15, torch 2.15.0 (MPS), coremltools 9.0, torchvision 0.29.2, timm 1.0.31, ml-depth-pro-2 commit 919ff37a3489b3597d4aa2e8fa6e7e95913e0c5a.
  • Phase 2 (MLX): MLX 0.32.4 (mlx-metal 0.32.4), depth_pro_2_mlx 0.1.0, from Origin commit 3dba66439173e51abdb791d4ccf0406dc23794ac.
  • The 4-bit conversion used the same MLX port, with affine 4-bit weight-only quantization (group size 64) on encoder Linear layers.

Timing protocol

Timing follows Apple's official demo/README.md protocol for the published MacBook Pro M4 Max table.

  • W=1 is single-frame timing. W=4 is unbuffered streaming.
  • The first prediction (which compiles for the device) is dropped, as are the next window + 2 warmup predictions.
  • The reported figure is the median milliseconds per frame of the remaining timed predictions (20 timed frames, or 10 if the median of the first three exceeded 250 ms/frame). FPS is 1000 / median_ms.
  • Preprocessing and rendering are not included.
  • Core ML used CPU_AND_GPU, matching the demo's .cpuAndGPU. Official W=4 packages came from apple/DepthPro2-coreml. W=1 packages were exported locally with python -m demo.export stream -w 1.
  • Matrix timings pass a fixed focal length, so they do not include the FOV estimator. End-to-end photo and video runs estimate focal length when EXIF is absent.
  • Inputs for the speed matrix are panning cover-crops of Apple's official assets/example.jpg. End-to-end quality runs use that photo at native 1920×1440 and the official example video (61 decoded frames).

The M4 Max columns in the tables below are Apple's published numbers from that README, measured on a 64GB MacBook Pro M4 Max. They are not from this Mac mini.

Headline results

W=1, H+ at 720p (ms/frame)

Runtime ms/frame
MLX BF16, this M5 Pro 472.7
Core ML, this M5 Pro 494.8
PyTorch MPS BF16, this M5 Pro 621.7
Apple published, M4 Max (different machine) 826

W=4 unbuffered streaming at 720p (ms/frame)

MLX BF16 was fastest on S (60.3), L (353.6), and H+ (629.1, versus 1,081.8 in Core ML). On B, PyTorch MPS was 135.4 ms and MLX BF16 was 139.4 ms.

Core ML still leads most single-frame cells

Core ML W=1 S ran at 149 fps at 360p (6.7 ms) and 31 fps at 720p (31.5 ms).

FOV head at 896×512

Runtime median ms FPS
MLX FP16 15.61 64.1
MLX BF16 15.66 63.9
Core ML 27.74 36.1

Parity (MLX vs Phase 1 PyTorch MPS references)

On the official photo, MLX BF16 inverse-depth Pearson r was at least 0.99995 for every size (S 0.999958, B 0.999976, L 0.999994, H+ 0.999996). On BF16 video, r was at least 0.99998.

mx.compile on S at 1280×720 W=1 was a modest 1.07× (BF16, 43.5 → 40.7 ms) to 1.08× (FP16, 51.9 → 48.2 ms).

W=1 S FPS by runtime

W=1 H+ ms by runtime

W=4 720p ms by size

FOV head ms by runtime

MLX vs PyTorch parity

Full speed matrices, including all five resolutions and Apple's published M4 Max figures, are in phase1/REPORT.md and phase2/REPORT_PHASE2.md, and in merged_3way_matrix.csv.

4-bit MLX conversion

I quantized the MLX port's encoder Linear layers to 4-bit (group size 64, affine) and checked them against the same Phase 1 photo references. The inclusion bar was no NaNs and inverse-depth Pearson r ≥ 0.999.

Size 4-bit Pearson r BF16 Pearson r 4-bit median rel. error NaNs Uploaded
S 0.99499 0.99996 0.0489 0 no
B 0.99653 0.99998 0.1552 0 no
L 0.99959 0.99999 0.0123 0 yes
H+ 0.99986 1.00000 0.0311 0 yes

Only L and H+ met the bar. Those sizes are in pratikgx/Depth-Pro-2-MLX-4bit (272.1 MiB and 562.6 MiB). S and B were converted with the same recipe and are not in that repo. The FOV estimator was kept at BF16/FP32; its 0.581 px difference versus the PyTorch reference is the same as MLX BF16.

4-bit vs BF16 parity

Caveats

  • MLX FP16 produced all-NaN depth maps on some cells (the L photo at 1920×1440, and buffered W=16 video for B and H+). Use BF16 for full precision, not FP16. Timing for those cells is still valid. S was fine in FP16 on the photo and video, and the H+ photo was fine in FP16.
  • Three optional FP16 video cells were skipped to stay inside the bench window: e2e_mp4_b_mlx_fp16_stream_w4, e2e_mp4_l_mlx_fp16_buffered_w16_o4, and e2e_mp4_l_mlx_fp16_stream_w4. Every required Table 1, FOV, photo, and BF16 video cell completed.
  • Buffered-video peak Metal memory readings are not trustworthy as per-cell maxima. Some back-to-back cells reported identical peaks, including values above the 64 GB of this machine. Prefer Table 1, photo, and unbuffered-stream peaks.
  • The 4-bit quality numbers come from a small test set (the official photo, plus a 1280×720 cover-crop compared to a same-session BF16 run). S and B missed the 0.999 Pearson bar, so only L and H+ were uploaded.
  • Video parity against Phase 1 should be read as approximate, because those video references were stored as float16.
  • Apple's M4 Max figures are published numbers from the official demo README, measured on a different machine.

Files

  • results_phase1.csv, results_phase2.csv, merged_3way_matrix.csv
  • phase1/ Phase 1 report, summary JSON, per-cell JSON, software specs, and the Core ML / PyTorch script
  • phase2/ Phase 2 report, summary JSON, per-cell JSON, software specs, the MLX script, and charts a–h
  • mlx-4bit/results_4bit.json and the 4-bit check script
  • charts/ Phase 2 charts a–h plus the 4-bit parity chart

This dataset does not include model weights or the large raw depth maps and videos.

Licenses and credit

Measurements and this card: CC BY 4.0, attribution Pratik Gandhi.

Apple Machine Learning Research Model is licensed under the Apple Machine Learning Research Model License Agreement. Depth Pro 2 is for research use only. No original or converted model weights are redistributed here. The 4-bit conversion lives in a separate model repo and carries that same license.

Credit to Apple and the Depth Pro 2 authors, and to @SpaceXSI for the MLX port, which was built with a coding agent.

Write-up on GitHub

How I ran this, what I measured, and the caveats: Apple Depth Pro 2 (MLX). It's part of Local AI field notes, where I log every model I run on my Mac mini.

Write-up on GitHub

How I ran this, what broke, and the scripts: Apple Depth Pro 2 (Core ML, PyTorch, MLX). It's part of Local AI field notes, where I log every model I run on my Mac mini.

Downloads last month
17