MoGe
Microsoft's MoGe: depth, 3D points, normals and focal length from one image.
stable Fork of microsoft/MoGe@74fbce0 · backends: portable, kernels
- Paper MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details arXiv · first released 3 Jul 2025
Paper 2025 Added in 0.1.0 Evaluated with 0.2.0
Use it
pip install fastforks Lean install with only what this model needs: pip install fastforks-core
from fastforks import depth
model = depth.load("moge-2-vits-normal") # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.depth # also: points, intrinsics, mask, normals
result.backend # what ran: mode, device, dtype, kernels fastforks bench moge-2-vits-normal Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.
Speed and memory
Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 162 | 0.51× | 1,491 MiB/1,628 MiB | 0.4 | 0.22 |
| Original fp16 | fp16 | 82.6 | – | 1,193 MiB/1,268 MiB | 0.9 | 1.50 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 61.6 | 1.3× | 1,007 MiB/1,172 MiB | 1.2 | 0.06 |
| Kernels | fp16 | 33.3 | 2.5× | 932 MiB/1,048 MiB | 0.2 | 0.03 |
| Portable + torch.compile | 43.9 | 1.9× | 836 MiB/992 MiB | 124 (build) | – | |
| TensorRT fp16 | fp16 | 24.5 | 3.4× | 620 MiBenginenot comparable | 131 (build) | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-05 to 2026-11-15. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two
times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.
Other devices
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14 | Original | fp32 | 516 | 2,849 MiB process | matches | 994x1482 |
| Portable | fp16 | 857 | 1,679 MiB process | matches | 994x1482 | |
| Portable | fp32 | 361 | 2,634 MiB process | matches | 994x1482 | |
| Kernels | fp32 | 362 | 2,704 MiB process | matches | 994x1482 | |
| NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 75.2 | 987 MiB/1,132 MiB | – | 994x1482 |
| Kernels | fp16 | 44.0 | 805 MiB/868 MiB | – | 994x1482 | |
| NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 24.9 | 987 MiB/1,132 MiB | – | 994x1482 |
| Kernels | fp16 | 18.0 | 805 MiB/868 MiB | – | 994x1482 | |
| NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 56.7 | 987 MiB/1,132 MiB | – | 994x1482 |
| Kernels | fp16 | 35.6 | 805 MiB/868 MiB | – | 994x1482 | |
| NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 26.5 | 987 MiB/1,132 MiB | – | 994x1482 |
| Kernels | fp16 | 22.1 | 805 MiB/868 MiB | – | 994x1482 | |
| NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 15.4 | 1,011 MiB/1,132 MiB | – | 994x1482 |
| Kernels | fp16 | 11.0 | 829 MiB/908 MiB | – | 994x1482 |
CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.
Licence and weights
Commercial use allowed
- Code
- MIT LICENSE
- Weights
- MIT source
- Training data
- Mixture of public synthetic and real RGB-D / multi-view datasets per the MoGe-2 and MoGe-3 papers
Our reading of the licences, not legal advice.
Weights
moge-2-vits-normal/model.pt 141 MB· hosted on Hugging Facefastforks downloads and checks these files on first use (fastforks download moge-2-vits-normal).
Accuracy
The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.
NYUv2 test
AbsRel / delta1
| Run | dtype | AbsRel ↓ | δ1 ↑ | Δ AbsRel vs fp32 original | images | ms / image |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.0484 | 0.9757 | reference | 654 | 163 |
| Original fp16 | fp16 | 0.0484 | 0.9757 | ±0.0000 | 654 | 82.5 |
| Portable fp32 | fp32 | 0.0484 | 0.9757 | ±0.0000 | 654 | 137 |
| Portable | fp16 | 0.0484 | 0.9757 | ±0.0000 | 654 | 63.5 |
| Kernels | fp16 | 0.0484 | 0.9757 | ±0.0000 | 654 | 31.5 |
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6d3bab64, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.
iBims-1
AbsRel / delta1
| Run | dtype | AbsRel ↓ | δ1 ↑ | Δ AbsRel vs fp32 original | images | ms / image |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.0430 | 0.9793 | reference | 100 | 161 |
| Original fp16 | fp16 | 0.0430 | 0.9793 | ±0.0000 | 100 | 81.3 |
| Portable fp32 | fp32 | 0.0430 | 0.9793 | ±0.0000 | 100 | 136 |
| Portable | fp16 | 0.0430 | 0.9793 | ±0.0000 | 100 | 62.8 |
| Kernels | fp16 | 0.0430 | 0.9793 | ±0.0000 | 100 | 32.7 |
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6d3bab64, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.
ETH3D two-view train left
AbsRel / delta1training data not stated: possibly seen
| Run | dtype | AbsRel ↓ | δ1 ↑ | Δ AbsRel vs fp32 original | images | ms / image |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.0700 | 0.9412 | reference | 27 | 162 |
| Original fp16 | fp16 | 0.0700 | 0.9411 | ±0.0000 | 27 | 82.3 |
| Portable fp32 | not measured | |||||
| Portable | fp16 | 0.0700 | 0.9411 | ±0.0000 | 27 | 63.9 |
| Kernels | fp16 | 0.0701 | 0.9412 | −0.0001 | 27 | 71.0 |
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.
Middlebury 2014 trainingH left
AbsRel / delta1training data not stated: possibly seen
| Run | dtype | AbsRel ↓ | δ1 ↑ | Δ AbsRel vs fp32 original | images | ms / image |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.0356 | 0.9947 | reference | 15 | 163 |
| Original fp16 | fp16 | 0.0356 | 0.9947 | ±0.0000 | 15 | 83.5 |
| Portable fp32 | not measured | |||||
| Portable | fp16 | 0.0356 | 0.9947 | ±0.0000 | 15 | 64.0 |
| Kernels | fp16 | 0.0356 | 0.9947 | ±0.0000 | 15 | 102 |
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 939adfe8, 2026-11-06.
NYUv2 test normals
mean / median angle deg, % < 11.25 / < 30 deg
| Run | dtype | ang_mean ↑ | ang_median ↑ | a11.25 ↑ | a30 ↑ | images | ms / image |
|---|---|---|---|---|---|---|---|
| Original fp32 | not measured | ||||||
| Original fp16 | not measured | ||||||
| Portable fp32 | fp32 | 15.87 | 9.25 | 58.32 | 83.97 | 654 | 164 |
| Portable | fp16 | 15.87 | 9.25 | 58.32 | 83.97 | 654 | 80.7 |
| Kernels | fp16 | 15.87 | 9.25 | 58.32 | 83.97 | 654 | 40.3 |
The original code was not run on this dataset, so the table cannot show the distance to it.
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Normals: angle to the ground-truth normal over the pixels with ground truth (Marigold's evaluation sets), mean over images. Measured on RTX 5060 Ti, torch 2.14.1+cu130, fastforks 4e890760, 2026-11-12.
iBims-1 normals
mean / median angle deg, % < 11.25 / < 30 deg
| Run | dtype | ang_mean ↑ | ang_median ↑ | a11.25 ↑ | a30 ↑ | images | ms / image |
|---|---|---|---|---|---|---|---|
| Original fp32 | not measured | ||||||
| Original fp16 | not measured | ||||||
| Portable fp32 | fp32 | 15.95 | 7.19 | 66.56 | 83.74 | 100 | 163 |
| Portable | fp16 | 15.95 | 7.19 | 66.56 | 83.74 | 100 | 79.5 |
| Kernels | fp16 | 15.95 | 7.19 | 66.56 | 83.74 | 100 | 39.3 |
The original code was not run on this dataset, so the table cannot show the distance to it.
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Measured on RTX 5060 Ti, torch 2.14.1+cu130, fastforks 4e890760, 2026-11-12.
Closeness to the original
Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. CPU, float32: max |diff| <= 0.005 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 1.5 % and p99 <= 1.5 % of the reference's median |value| (absolute error of the output depth).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.005 (relative) | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.015 (relative) | 0.00116 (relative) |
| GPU default dtypes, p99 |diff| | ≤ 0.015 (relative) | 0.0013 (relative) |
Recorded for moge-2-vits-normal on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-07, with these kernels: dpt_fusedfastbbsparse_conv
Details
MoGe-2 / MoGe-3 (Wang et al., Microsoft): monocular metric geometry; a DINOv2 ViT with conv heads predicts an affine point map, a validity mask, normals and a metric scale, and the focal length and shift are recovered from the points (MoGe-3 adds a sparse 3-D refiner).
- Input / output
- Input: one RGB image. Output: a depth or inverse-depth map.
What the numbers mean, and which models need the camera focal length
kind(a card field,result.kind) says what the numbers mean:metric= metres;scale_invariant,affine_invariant,affine_invariant_log(log depth) andaffine_invariant_inverseare relative (unknown scale or shift, not metres). Metric models that use the focal length takemodel(image, intrinsics=K)(K[3, 3], pixels of the input):metric3d-vit-largeandda3metric-largeneed it for correct metres (da3metric-large: metres = output x focal / 300; withoutintrinsics=a 300 px focal length is assumed, wrong for a real camera).unidepth-v2-vitl14,unik3d-vitl,depth-proand MoGe estimate the camera themselves (givenintrinsics=replacefx);zoedepth-nkand the relative models (da3mono-largetoo) take none. Normals-only models (kindnone) returnresult.normals[B, H, W, 3](unit vectors, camera frame: x right, y down, z forward). - Input size
- any H x W >= 8, resized or padded inside the model (ViT patch 14); tested up to 3000x4000 (1,257 MiB allocated)
- Options
resolution_level=90 to 9· changes the outputnormals=false· changes the output
- Outputs
depth,points,intrinsics,mask,normals- Checkpoint
arch="vits",version=2- CUDA kernels
fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbonesdpt_fusedfused fp16 matrix multiplies and convolutions for DINOv2 backbones and DPT depth headssparse_convMoGe-3: sparse 3D convolution
- Apple (MPS)
sparse_metalMetal: sparse 3D convolution. fp32, fp16
Limitations
- MoGe-3 compiles the dense network and leaves sparse voxel refinement eager. Its 100-image iBims-1 comparison used the default CUDA float16 mode.
- At 994 x 1482, the portable and compiled MoGe-3 timings use the same synthetic image. The original timings use the Middlebury Motorcycle image.
Full reference, generated from the repository
Microsoft's MoGe: depth, 3D points, normals and focal length from one image.
Commercial use: yes.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
moge-2-vits-normal | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 162 82.6 61.6 33.3 43.9 24.5 | 1491 / 1628 1193 / 1268 1007 / 1172 932 / 1048 836 / 992 768 / 852 |
moge-2-vitb-normal | synthetic (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 237 102 83.9 40.7 60.9 35.5 | 1773 / 1876 1469 / 1596 1163 / 1312 1093 / 1208 972 / 1172 1044 / 1092 |
moge-2-vitl-normal | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 501 171 144 73.4 124 75.6 | 2646 / 2818 2339 / 2438 1661 / 1836 1498 / 1610 1411 / 1634 1882 / 1936 |
moge-3-vitl | Middlebury 2014 half (994x1482) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 843 486 334 149 311 planned for 0.2.1 | 2812 / 2990 2505 / 2630 1678 / 1826 1626 / 1762 1500 / 1724 – |
Details
- Original code: https://github.com/microsoft/MoGe at
74fbce0 - Code licence: MIT (permissive); files:
LICENSE - Weights licence: MIT
- Licence evidence: https://github.com/microsoft/MoGe/blob/main/README.md (License section) and the Hugging Face model cards (license: mit)
- Training data: Mixture of public synthetic and real RGB-D / multi-view datasets per the MoGe-2 and MoGe-3 papers
- Options:
resolution_level= 9 (0 to 9),normals= False - Kernel libraries:
fastbb,dpt_fused,sparse_conv - Variants:
moge-2-vits-normal,moge-2-vitb-normal,moge-2-vitl-normal,moge-3-vitl - Not in this release (planned for 0.2.1):
moge-3-vitlONNX / TensorRT export: torch.export fails (TorchExportError, step 1/3 of the ONNX export) - Outputs: depth, points, intrinsics, mask, normals; kind:
metric - Input size:
moge-2-vits-normal,moge-2-vitl-normal,moge-3-vitl: any H x W >= 8, resized or padded inside the model (ViT patch 14)moge-2-vitb-normal: any H x W >= 1, resized or padded inside the model (ViT patch 14)
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
moge-2-vits-normal3000x4000 (1.2 GiB peak allocated, 1.4 reserved);moge-2-vitb-normal3000x4000 (1.3 GiB peak allocated, 1.4 reserved);moge-2-vitl-normal3000x4000 (2.4 GiB peak allocated, 2.5 reserved);moge-3-vitl3000x4000 (2.6 GiB peak allocated, 3.0 reserved) moge-2-vits-normal994x1482, first call: portable + torch.compile 124 s (cold Inductor cache), TensorRT fp16 131 s (ONNX + engine).moge-2-vitb-normal994x1482, first call: portable + torch.compile 133 s (cold Inductor cache), TensorRT fp16 127 s (ONNX + engine).moge-2-vitl-normal994x1482, first call: portable + torch.compile 152 s (cold Inductor cache), TensorRT fp16 173 s (ONNX + engine).moge-3-vitl994x1482, first call: portable + torch.compile 136 s (cold Inductor cache).moge-2-vits-normalload / first call: original 0.88 s / 1.501 s, portable 1.18 s / 0.298 s (518x770), kernels 0.21 s / 0.438 s (518x770)moge-2-vitb-normalload / first call: original 1.26 s / 0.239 s, portable 0.49 s / 0.397 s (518x770), kernels 1.24 s / 0.285 s (518x770)moge-2-vitl-normalload / first call: original 7.24 s / 0.171 s, portable 1.37 s / 0.144 s, kernels 0.63† s / 0.131 smoge-3-vitlload / first call: original 10.6 s / 2.918 s, portable 2.31 s / 0.69 s (518x770), kernels 1.16 s / 0.742 s (518x770)- † measured in the speed sweep after the portable load (or an earlier size) in the same process: file cache and kernels are warm, so the time is lower than a fresh start.
- Protocol,
moge-2-vits-normal,moge-2-vitb-normal,moge-3-vitl: first call, 3 warm-up, peak reset, median of 3 rounds x median of 5 - Protocol,
moge-2-vitl-normal: median of 5 after 3 warm-up, uint8 CUDA input (synthetic for bench_inputs sizes), card defaults; gc before each load; peak = max over the first call and the timed calls
Measured
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Normals: angle to the ground-truth normal over the pixels with ground truth (Marigold's evaluation sets), mean over images. Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | original fp16 | portable | portable fp32 | kernels | portable + torch.compile | note |
|---|---|---|---|---|---|---|---|---|
moge-2-vits-normal | NYUv2 test, AbsRel / delta1 | 0.0484 / 0.9757 | 0.0484 / 0.9757 | 0.0484 / 0.9757 | 0.0484 / 0.9757 | 0.0484 / 0.9757 | ||
moge-2-vits-normal | iBims-1, AbsRel / delta1 | 0.0430 / 0.9793 | 0.0430 / 0.9793 | 0.0430 / 0.9793 | 0.0430 / 0.9793 | 0.0430 / 0.9793 | ||
moge-2-vits-normal | ETH3D two-view train left, AbsRel / delta1 | 0.0700 / 0.9412 | 0.0700 / 0.9411 | 0.0700 / 0.9411 | 0.0701 / 0.9412 | possibly in the training data | ||
moge-2-vits-normal | Middlebury 2014 trainingH left, AbsRel / delta1 | 0.0356 / 0.9947 | 0.0356 / 0.9947 | 0.0356 / 0.9947 | 0.0356 / 0.9947 | possibly in the training data | ||
moge-2-vits-normal | NYUv2 test normals, mean / median angle deg, % < 11.25 / < 30 deg | 15.87 / 9.25 / 58.32 / 83.97 | 15.87 / 9.25 / 58.32 / 83.97 | 15.87 / 9.25 / 58.32 / 83.97 | ||||
moge-2-vits-normal | iBims-1 normals, mean / median angle deg, % < 11.25 / < 30 deg | 15.95 / 7.19 / 66.56 / 83.74 | 15.95 / 7.19 / 66.56 / 83.74 | 15.95 / 7.19 / 66.56 / 83.74 | ||||
moge-2-vitb-normal | NYUv2 test, AbsRel / delta1 | 0.0404 / 0.9812 | 0.0404 / 0.9812 | 0.0404 / 0.9812 | 0.0404 / 0.9812 | 0.0404 / 0.9812 | ||
moge-2-vitb-normal | iBims-1, AbsRel / delta1 | 0.0337 / 0.9855 | 0.0337 / 0.9854 | 0.0337 / 0.9854 | 0.0337 / 0.9855 | 0.0337 / 0.9854 | ||
moge-2-vitb-normal | ETH3D two-view train left, AbsRel / delta1 | 0.0667 / 0.9492 | 0.0667 / 0.9492 | 0.0667 / 0.9492 | 0.0663 / 0.9496 | possibly in the training data | ||
moge-2-vitb-normal | Middlebury 2014 trainingH left, AbsRel / delta1 | 0.0260 / 0.9962 | 0.0260 / 0.9962 | 0.0260 / 0.9962 | 0.0261 / 0.9962 | possibly in the training data | ||
moge-2-vitb-normal | NYUv2 test normals, mean / median angle deg, % < 11.25 / < 30 deg | 15.25 / 8.61 / 60.32 / 84.90 | 15.25 / 8.61 / 60.32 / 84.90 | 15.25 / 8.61 / 60.33 / 84.90 | ||||
moge-2-vitb-normal | iBims-1 normals, mean / median angle deg, % < 11.25 / < 30 deg | 15.28 / 6.98 / 68.57 / 85.01 | 15.28 / 6.97 / 68.57 / 85.01 | 15.28 / 6.98 / 68.56 / 85.01 | ||||
moge-2-vitl-normal | NYUv2 test, AbsRel / delta1 | 0.0357 / 0.9836 | 0.0357 / 0.9836 | 0.0357 / 0.9836 | 0.0357 / 0.9836 | 0.0357 / 0.9836 | ||
moge-2-vitl-normal | iBims-1, AbsRel / delta1 | 0.0261 / 0.9892 | 0.0261 / 0.9892 | 0.0261 / 0.9892 | 0.0261 / 0.9892 | 0.0261 / 0.9892 | ||
moge-2-vitl-normal | ETH3D two-view train left, AbsRel / delta1 | 0.0518 / 0.9789 | 0.0518 / 0.9789 | 0.0519 / 0.9789 | 0.0518 / 0.9790 | possibly in the training data | ||
moge-2-vitl-normal | Middlebury 2014 trainingH left, AbsRel / delta1 | 0.0235 / 0.9972 | 0.0235 / 0.9972 | 0.0235 / 0.9972 | 0.0235 / 0.9972 | possibly in the training data | ||
moge-2-vitl-normal | NYUv2 test normals, mean / median angle deg, % < 11.25 / < 30 deg | 14.81 / 8.07 / 61.96 / 85.54 | 14.81 / 8.07 / 61.96 / 85.54 | 14.81 / 8.07 / 61.96 / 85.55 | ||||
moge-2-vitl-normal | iBims-1 normals, mean / median angle deg, % < 11.25 / < 30 deg | 14.70 / 6.49 / 70.49 / 85.80 | 14.70 / 6.49 / 70.49 / 85.80 | 14.69 / 6.49 / 70.49 / 85.80 | ||||
moge-3-vitl | NYUv2 test, AbsRel / delta1 | 0.0356 / 0.9836 | 0.0356 / 0.9836 | 0.0356 / 0.9836 | 0.0356 / 0.9836 | 0.0356 / 0.9836 | ||
moge-3-vitl | iBims-1, AbsRel / delta1 | 0.0261 / 0.9884 | 0.0261 / 0.9884 | 0.0261 / 0.9884 | 0.0261 / 0.9884 | 0.0261 / 0.9883 | 0.0261 / 0.9884 | |
moge-3-vitl | ETH3D two-view train left, AbsRel / delta1 | 0.0496 / 0.9847 | 0.0496 / 0.9847 | 0.0496 / 0.9847 | 0.0496 / 0.9847 | possibly in the training data | ||
moge-3-vitl | Middlebury 2014 trainingH left, AbsRel / delta1 | 0.0235 / 0.9970 | 0.0235 / 0.9970 | 0.0235 / 0.9970 | 0.0235 / 0.9970 | possibly in the training data |
Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 24f9c9c8 / 3d00db1d / 4e890760 / 939adfe8, 2026-11-06 to 2026-11-12; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 24f9c9c8 / 4e890760 / 6d3bab64, 2026-11-12 to 2026-11-17; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 3d00db1d / 4e890760 / 939adfe8 / f3195fb2, 2026-11-06 to 2026-11-12; compile: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks ed52b189, 2026-11-18.
Parity
CPU, float32: max |diff| <= 0.005 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 1.5 % and p99 <= 1.5 % of the reference's median |value| (absolute error of the output depth).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
moge-2-vits-normal | 0.00116 / 0.0013 | RTX 5060 Ti, torch 2.14.1+cu130 | dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120, sparse_conv@2feb2d5e0877:sm_120 |
moge-2-vitb-normal | 0.02172 / 0.02188 | RTX 5060 Ti, torch 2.14.1+cu130 | dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120, sparse_conv@2feb2d5e0877:sm_120 |
moge-2-vitl-normal | 0.00382 / 0.00394 | RTX 5060 Ti, torch 2.14.1+cu130 | dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120, sparse_conv@2feb2d5e0877:sm_120 |
moge-3-vitl | 0.00421 / 0.0049 | RTX 5060 Ti, torch 2.14.1+cu130 | dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120, sparse_conv@2feb2d5e0877:sm_120 |
Limitations
- MoGe-3 compiles the dense network and leaves sparse voxel refinement eager. Its 100-image iBims-1 comparison used the default CUDA float16 mode.
- At 994 x 1482, the portable and compiled MoGe-3 timings use the same synthetic image. The original timings use the Middlebury Motorcycle image.
Example output
One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).
Input: A cup of coffee, Rachel Michetti (CC0).