OMNI-DC
OMNI-DC: metric depth completion for any sparse pattern; it predicts depth gradients at three scales and integrates them with the sparse points, helped by Depth Anything V2 Large. BSD-3 code, weights non-commercial (they include Depth Anything V2 Large).
experimental Fork of princeton-vl/OMNI-DC@a5ecd35 · backends: portable, kernels
- Paper OMNI-DC: Highly Robust Depth Completion with Multiresolution Depth Integration ICCV 2025 · first released 28 Nov 2024
Checkpoint omni-dc
Paper 2024 Added in 0.2.0 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("omni-dc") # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.depth
result.backend # what ran: mode, device, dtype, kernels The weights are non-commercial: fastforks downloads them only with FASTFORKS_NONCOMMERCIAL=1 set.
fastforks bench omni-dc 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.
Input: synthetic, 994x1482 (the task’s standard size)
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | not measured | – | – | – | – |
| Original (own default)upstream OGNIDC.forward (demo.sh arguments), float32, 16-padded | fp32 | 875 | – | 5,966 MiB/6,732 MiB | 9.5 | 2.04 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 554 | 1.6× | 4,958 MiB/6,304 MiB | 2.5 | 1.06 |
| Kernels | fp16 | 510 | 1.7× | 4,981 MiB/6,030 MiB | 2.4 | 0.94 |
| Portable + torch.compile | 411 | 2.1× | 4,201 MiB/5,932 MiB | 137 (build) | – | |
| TensorRTplanned for 0.2.1 | fails | – | – | – | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-15 to 2026-11-16. 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, 2026-11-17 | Portable | fp16 | 2408 | 10,081 MiB process | skip | 994x1482 |
| Portable | fp32 | 2349 | 12.0 GiB process | skip | 994x1482 | |
| Kernels | fp16 | not implemented | – | 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
Non-commercial only
- Code
- BSD-3-Clause LICENSE
- Weights
- CC-BY-NC-4.0 source
- Training data
- Synthetic RGB-D datasets with virtual sparse patterns per the paper (arXiv 2411.19278)
- weights: CC-BY-NC-4.0
Our reading of the licences, not legal advice.
Weights
model.safetensors 1.7 GB· hosted on Hugging Facefastforks downloads and checks these files on first use (fastforks download omni-dc). Non-commercial weights need FASTFORKS_NONCOMMERCIAL=1.
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 (every 4th: 164 images) with a 192x256 depth prompt
AbsRel / delta1 (metric)
| Run | dtype | met_absrel ↑ | met_d1 ↑ | Δ met_absrel vs fp32 original | images | ms / image |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.0231 | 0.9966 | reference | 164 | 327 |
| Original fp16 | not measured | |||||
| Portable fp32 | not measured | |||||
| Portable | fp16 | 0.0231 | 0.9966 | ±0.0000 | 164 | 165 |
| Kernels | fp16 | 0.0232 | 0.9966 | +0.0001 | 164 | 120 |
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Prompt depth models: the prompt is a 192x256 low-resolution depth built from the ground truth (mean of the valid pixels per cell, holes filled with the nearest cell), the output is metric depth compared without any scale fit; AbsRel / delta1 over the pixels with ground truth, mean over images. The prompt comes from the ground truth, so the numbers are not comparable with the paper's (ARKitScenes / ScanNet++ with a LiDAR prompt, L1 in mm): they compare our modes with the original. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-16; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 68b2f345, 2026-11-16.
iBims-1 (100 images) with a 192x256 depth prompt
AbsRel / delta1 (metric)
| Run | dtype | met_absrel ↑ | met_d1 ↑ | Δ met_absrel vs fp32 original | images | ms / image |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.0253 | 0.9910 | reference | 100 | 336 |
| Original fp16 | not measured | |||||
| Portable fp32 | not measured | |||||
| Portable | fp16 | 0.0254 | 0.9910 | +0.0001 | 100 | 175 |
| Kernels | fp16 | 0.0254 | 0.9910 | +0.0001 | 100 | 128 |
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Prompt depth models: the prompt is a 192x256 low-resolution depth built from the ground truth (mean of the valid pixels per cell, holes filled with the nearest cell), the output is metric depth compared without any scale fit; AbsRel / delta1 over the pixels with ground truth, mean over images. The prompt comes from the ground truth, so the numbers are not comparable with the paper's (ARKitScenes / ScanNet++ with a LiDAR prompt, L1 in mm): they compare our modes with the original. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-16; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 68b2f345, 2026-11-16.
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 <= 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) | 0.00021 (relative) |
| GPU default dtypes, p99 |diff| | ≤ 0.05 (relative) | 0.00144 (relative) |
Recorded for omni-dc on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-14, with these kernels: dpt_fusedfastbb
Details
OMNI-DC (ICCV 2025): zero-shot depth completion for any sparse pattern; an RGB-D network predicts log-depth gradients at three resolutions, a conjugate-gradient solve integrates them with the sparse depth, DySPN refines the result. Uses Depth Anything V2 Large as an extra input.
- 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 >= 1, resized or padded inside the model (ViT patch 16); tested up to 1024x2048 (6,544 MiB allocated)
- Outputs
depth- 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 heads
- Apple (MPS)
- fp32, fp16
Limitations
- The depth map is required: call the model with `prompt_depth=` a metric map in metres at the image size (another size is resized nearest; missing points as 0 / NaN / inf).
- The weights include the frozen Depth Anything V2 Large (CC-BY-NC-4.0), so they are labelled non-commercial although the code and the Hugging Face card are BSD-3-Clause.
- The integration is an iterative solve with a data-dependent step count: no ONNX / TensorRT export.
Full reference, generated from the repository
OMNI-DC: metric depth completion for any sparse pattern; it predicts depth gradients at three scales and integrates them with the sparse points, helped by Depth Anything V2 Large. BSD-3 code, weights non-commercial (they include Depth Anything V2 Large).
Commercial use: no (weights: CC-BY-NC-4.0).
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
omni-dc | synthetic (994x1482) | original portable kernels portable + torch.compile TensorRT fp16 | 875 554 510 411 planned for 0.2.1 | 5966 / 6732 4958 / 6304 4981 / 6030 4201 / 5932 – |
Details
- Original code: https://github.com/princeton-vl/OMNI-DC at
a5ecd35 - Code licence: BSD-3-Clause (permissive); files:
LICENSE - Weights licence: CC-BY-NC-4.0
- Licence evidence: https://github.com/princeton-vl/OMNI-DC/blob/main/LICENSE (BSD-3-Clause, also on the Hugging Face card zuoym15/OMNI-DC); the checkpoint's depth_module is the frozen Depth Anything V2 Large (upstream loads depth_anything_v2_vitl.pth, CC-BY-NC-4.0)
- Training data: Synthetic RGB-D datasets with virtual sparse patterns per the paper (arXiv 2411.19278)
- Kernel libraries:
fastbb,dpt_fused - Variants:
omni-dc - Not in this release (planned for 0.2.1):
omni-dcONNX / TensorRT export: torch.export fails (TorchExportError, step 1/3 of the ONNX export) - Outputs: depth; kind:
metric - Input size: any H x W >= 1, resized or padded inside the model (ViT patch 16)
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
omni-dc1024x2048 (6.4 GiB peak allocated, 9.2 reserved) omni-dc994x1482, first call: portable + torch.compile 137 s (cold Inductor cache).omni-dcoriginal 994x1482: upstream OGNIDC.forward (demo.sh arguments), float32, 16-paddedomni-dcload / first call: original 9.52 s / 2.035 s, portable 2.47 s / 1.062 s, kernels 2.44 s / 0.936 s- Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5
Measured
AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Prompt depth models: the prompt is a 192x256 low-resolution depth built from the ground truth (mean of the valid pixels per cell, holes filled with the nearest cell), the output is metric depth compared without any scale fit; AbsRel / delta1 over the pixels with ground truth, mean over images. The prompt comes from the ground truth, so the numbers are not comparable with the paper's (ARKitScenes / ScanNet++ with a LiDAR prompt, L1 in mm): they compare our modes with the original. Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | portable | kernels | note |
|---|---|---|---|---|---|
omni-dc | NYUv2 test (every 4th: 164 images) with a 192x256 depth prompt, AbsRel / delta1 (metric) | 0.0231 / 0.9966 | 0.0231 / 0.9966 | 0.0232 / 0.9966 | |
omni-dc | iBims-1 (100 images) with a 192x256 depth prompt, AbsRel / delta1 (metric) | 0.0253 / 0.9910 | 0.0254 / 0.9910 | 0.0254 / 0.9910 |
Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 68b2f345, 2026-11-16.
Parity
CPU, float32: max |diff| <= 0.005 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 1.5 % and p99 <= 5 % of the reference's median |value| (absolute error of the output depth).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
omni-dc | 0.00021 / 0.00144 | RTX 5060 Ti, torch 2.14.1+cu130 | dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120 |
Limitations
- The depth map is required: call the model with
prompt_depth=a metric map in metres at the image size (another size is resized nearest; missing points as 0 / NaN / inf). - The weights include the frozen Depth Anything V2 Large (CC-BY-NC-4.0), so they are labelled non-commercial although the code and the Hugging Face card are BSD-3-Clause.
- The integration is an iterative solve with a data-dependent step count: no ONNX / TensorRT export.
Example output
No example rendered for this checkpoint yet.