Hi3Dv1.0-mlx / validation.json
cguldogan's picture
Add Hi3D v1.0 MLX bundle (mlx-vlm hi3d)
195cfd9 verified
Raw History Blame Contribute Delete
3.87 kB
{
"mlx_version": "0.32.4",
"tensor_count": 1899,
"parameter_elements": 4312697600,
"weight_bytes": 10158060544,
"dtypes": {
"float32": 1151,
"float16": 748
},
"strict_load": {
"missing": 0,
"extra": 0,
"shape_mismatches": 0
},
"regression_tests": {
"class": "TestHi3D",
"passed": 14
},
"reference_comparison": {
"reference": "Original Hi3D PyTorch code (github 23dcc68, hub 705fc09) on CPU; spconv and flash-attn replaced by reference implementations; float32 unless noted",
"mlx_device": "GPU",
"reports": {
"sphere64_report": {
"conditioning": {
"max_abs": 5.714595317840576e-06,
"mean_abs": 2.006490236685449e-07,
"rel_l2": 8.875883576616418e-06
},
"sigmas": {
"max_abs": 5.960464477539063e-08,
"mean_abs": 4.584972675030048e-09,
"rel_l2": 2.0996074718879118e-08
},
"times": {
"max_abs": 5.960464477539063e-08,
"mean_abs": 4.76837158203125e-09,
"rel_l2": 2.0855193972367177e-08
},
"velocity_fp16_vs_fp32ref": {
"max_abs": 0.013125389814376831,
"mean_abs": 0.0010297759424406192,
"rel_l2": 0.0012895214320252043
},
"velocity_fp16_vs_fp16ref": {
"max_abs": 0.015636205673217773,
"mean_abs": 0.001322876022487012,
"rel_l2": 0.0016291599890619199
},
"ref_fp16_vs_ref_fp32": {
"max_abs": 0.016333699226379395,
"mean_abs": 0.0011320557014777956,
"rel_l2": 0.0013910033166357426
},
"velocity_fp32": {
"max_abs": 2.0503997802734375e-05,
"mean_abs": 1.8179455511446113e-06,
"rel_l2": 2.220910126120128e-06
},
"decoded_voxels": {
"ours": 1442341,
"reference": 1442339,
"common": 1442333
},
"decoded_sdf_sign": {
"max_abs": 28.49630832672119,
"mean_abs": 0.000985199632137544,
"rel_l2": 0.0035812783661513725
},
"decoded_feat": {
"max_abs": 0.3575434908270836,
"mean_abs": 2.4584798492266753e-05,
"rel_l2": 0.0023846487551618238
},
"mesh_faces_identical": true,
"mesh_vertices": {
"max_abs": 2.86102294921875e-06,
"mean_abs": 2.4002885922643986e-07,
"rel_l2": 5.207820769835567e-08
},
"filled_faces_identical": true
},
"sphere192_report": {
"conditioning": {
"max_abs": 5.714595317840576e-06,
"mean_abs": 2.006490236685449e-07,
"rel_l2": 8.875883576616418e-06
},
"sigmas": {
"max_abs": 0.0,
"mean_abs": 0.0,
"rel_l2": 0.0
},
"times": {
"max_abs": 0.0,
"mean_abs": 0.0,
"rel_l2": 0.0
},
"velocity_fp16_vs_fp32ref": {
"max_abs": 0.032924532890319824,
"mean_abs": 0.0010041114637135014,
"rel_l2": 0.001261798954717289
},
"velocity_fp32": {
"max_abs": 5.2988529205322266e-05,
"mean_abs": 2.1241101601153427e-06,
"rel_l2": 2.638211684786904e-06
}
}
}
},
"benchmark": {
"hardware": "Apple M4 Pro, 64 GiB unified memory",
"python": "3.12.11",
"input": "Hi3D assets/example.png (436x640 RGBA), coarse mesh from mlx-community/sam-3d-objects-bf16 (seed 0)",
"resolution": 1024,
"steps": 25,
"seed": 0,
"timings": {
"coarse_mesh": 77.6,
"load": 1.5,
"conditioning": 0.58,
"pass1_sampling": 189.08,
"pass1_decoding": 8.86,
"pass2_sampling": 1517.58,
"pass2_decoding": 44.96,
"postprocess": 8.67
},
"peak": 30.05,
"vertices": 495425,
"faces": 1000000
}
}