MoGe (fastforks)
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).
This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/microsoft/MoGe at 74fbce0.
Author weight repos: https://huggingface.co/Ruicheng/moge-2-vitl-normal, https://huggingface.co/Ruicheng/moge-2-vits-normal, https://huggingface.co/Ruicheng/moge-3-vitl.
Task: depth. Tier: experimental. Outputs: depth, points, intrinsics, mask, normals. Backends: portable, kernels.
Input: one RGB image. Output: a depth or inverse-depth map (metric for the models marked metric).
Variants
Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.
| variant | weights (author link) | SHA-256 | weights licence | commercial use |
|---|---|---|---|---|
moge-2-vitl-normal@1 |
model.pt | 280741fd09bc |
MIT | yes |
moge-2-vits-normal@1 |
model.pt | 79a16621928c |
MIT | yes |
moge-3-vitl@1 |
model.pt | 9b41b7b9f65a |
MIT | yes |
Licence
Original-code strings, recorded as they are (research-use package; check them before any other use). Code: MIT. Files: LICENSE.
Training data: Mixture of public synthetic and real RGB-D / multi-view datasets per the MoGe-2 and MoGe-3 papers.
Licence evidence: https://github.com/microsoft/MoGe/blob/main/README.md (License section) and the Hugging Face model cards (license: mit)
Input size
any H x W >= 8, resized or padded inside the model (ViT patch 14)
Tested up to (portable, fp16, RTX 5060 Ti 16 GB): moge-2-vitl-normal 3000x4000 (2.4 GiB peak allocated, 2.5 reserved); moge-2-vits-normal 3000x4000 (1.2 GiB peak allocated, 1.4 reserved); moge-3-vitl 3000x4000 (2.6 GiB peak allocated, 3.0 reserved)
Speed and memory
Depth, Middlebury 2014 half, 994x1482
| model | time (ms): original fp32 | original | portable | kernels | kernels speed-up: vs original | vs portable | peak memory (MiB, allocated / reserved): original | portable | kernels |
|---|---|---|---|---|---|---|---|---|---|
moge-2-vitl-normal |
502.6 | – | 144.5 | 73.43 | – | 2.0x | – | 1661 / 1836 | 1498 / 1610 |
moge-2-vits-normal |
161.5 | – | 64.19 | 29.73 | – | 2.2x | – | 1066 / 1336 | 884 / 1008 |
moge-3-vitl |
– | – | 904.91 | 507.19 | – | 1.8x | – | 2157 / 2424 | 1756 / 1886 |
Startup and first call
| model | load (s): original | portable | kernels | first call (s): original | portable | kernels |
|---|---|---|---|---|---|---|
moge-2-vitl-normal |
3.26 | 1.37 | 0.63† | 0.50 | 0.14 | 0.13 |
moge-2-vits-normal |
0.36 | 0.35 | 0.11† | 0.21 | 0.08 | 0.10 |
moge-3-vitl |
– | 1.47 | 0.69† | – | 0.95 | 0.65 |
Fixed-size speed and peak memory, all models
| model | mode | 640x480 | 1280x720 | 1920x1080 | 12 MP 4000x3000 |
|---|---|---|---|---|---|
moge-2-vitl-normal |
portable | 175.5 ms / 2275 / 2508 | 178.8 ms / 2286 / 2518 | 179.2 ms / 2299 / 2528 | 191.5 ms / 2410 / 2562 |
moge-2-vitl-normal |
kernels | 74.3 ms / 1435 / 1572 | 77.5 ms / 1444 / 1588 | 76.0 ms / 1459 / 1644 | 90.5 ms / 1755 / 2174 |
moge-2-vits-normal |
portable | 88.6 ms / 1123 / 1418 | 85.4 ms / 1132 / 1392 | 86.7 ms / 1146 / 1446 | 98.6 ms / 1257 / 1464 |
moge-2-vits-normal |
kernels | 33.9 ms / 818 / 970 | 33.1 ms / 826 / 976 | 31.0 ms / 839 / 1010 | 46.9 ms / 1144 / 1566 |
moge-3-vitl |
portable | 657.4 ms / 2580 / 3070 | 633.4 ms / 2577 / 2792 | 620.0 ms / 2588 / 2804 | 655.1 ms / 2701 / 3066 |
moge-3-vitl |
kernels | 386.2 ms / 1570 / 1836 | 374.2 ms / 1557 / 1840 | 362.3 ms / 1566 / 1890 | 390.6 ms / 1838 / 2300 |
Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).
Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the authors' code at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.
Accuracy
Parity with the original code
Every variant is checked against golden outputs of the original code at its pinned commit. 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|.
Usage
from fastforks import depth
model = depth.load("moge-2-vitl-normal") # portable PyTorch on the GPU if there is one, else the CPU
result = model("room.jpg")
result.depth # [H, W] float32 tensor; result.kind says what the maps mean
# models that take a camera: model("room.jpg", intrinsics=K), K [3, 3] in pixels
result.backend # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')
Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: moge-2-vitl-normal, moge-2-vits-normal, moge-3-vitl.
Kernels backend
depth.load("moge-2-vitl-normal", mode="kernels") runs our CUDA kernels (fastbb, dpt_fused, sparse_conv). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.