VGGT-Omega
VGGT-Omega (Meta): the successor of VGGT with a DINOv3 backbone, cameras and depth from many images. Non-commercial licence.
stable Fork of facebookresearch/vggt-omega@48b23c8 · backends: portable, kernels
- Paper VGGT-Ω CVPR 2026 · first released 14 May 2026
Checkpoint vggt-omega-1b-512
Paper 2026 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[nc]"
from fastforks import multiview
model = multiview.load("vggt-omega-1b-512") # kernels on a supported GPU, else portable PyTorch
result = model(images)
result.depth # also: confidence, intrinsics, cam_to_world
result.backend # what ran: mode, device, dtype, kernels The weights are non-commercial: fastforks downloads them only with FASTFORKS_NONCOMMERCIAL=1 set.
fastforks bench vggt-omega-1b-512 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, 336x518
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32float32 (the authors' autocast off) | fp32 | 2650 | 0.34× | 5,765 MiB/5,912 MiB | 12.3 | 2.64 |
| Original fp16original default VGGTOmega(autocast=True): bfloat16 aggregator, float32 heads | fp16 | 906 | – | 7,495 MiB/7,654 MiB | 16.3 | 1.10 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 726 | 1.2× | 3,421 MiB/3,518 MiB | 3.3 | 0.73 |
| Kernels | fp16 | 467 | 1.9× | 3,404 MiB/3,494 MiB | 5.2 | 0.75 |
| Portable + torch.compile | 589 | 1.5× | 3,082 MiB/3,154 MiB | 66 (build) | – | |
| TensorRT fp16 | fp16 | 377 | 2.4× | 2,994 MiBenginenot comparable | 187 (build) | – |
Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-03 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, 2026-11-12 | Original | fp32 | 7828 | 9,079 MiB process | matches | 336x518 |
| Portable | fp16 | 7242 | 8,602 MiB process | matches | 336x518 | |
| Portable | fp32 | 7938 | 8,221 MiB process | matches | 336x518 | |
| Kernels | fp16 | not implemented | – | 336x518 | ||
| NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 758 | 3,380 MiB/3,476 MiB | – | 336x518 |
| Kernels | fp16 | 599 | 3,380 MiB/3,494 MiB | – | 336x518 | |
| NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 199 | 3,380 MiB/3,476 MiB | – | 336x518 |
| Kernels | fp16 | 166 | 3,380 MiB/3,494 MiB | – | 336x518 | |
| NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 578 | 3,380 MiB/3,476 MiB | – | 336x518 |
| Kernels | fp16 | 447 | 3,380 MiB/3,434 MiB | – | 336x518 | |
| NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 210 | 3,380 MiB/3,476 MiB | – | 336x518 |
| Kernels | fp16 | 150 | 3,380 MiB/3,494 MiB | – | 336x518 | |
| NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 110 | 3,404 MiB/3,556 MiB | – | 336x518 |
| Kernels | fp16 | 87.6 | 3,404 MiB/3,534 MiB | – | 336x518 | |
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
- LicenseRef-FAIR-Noncommercial-Research LICENSE
- Weights
- LicenseRef-FAIR-Noncommercial-Research source
- Training data
- Per the repository README: a mixture of public multi-view datasets (uCO3D annotations released); DINOv3 initialisation. The 512 checkpoint is recommended for in-the-wild use, not benchmarking (vggt_omega_1b_416_reproduce.pt is the benchmark reference).
- code: LicenseRef-FAIR-Noncommercial-Research (noncommercial)
- weights: LicenseRef-FAIR-Noncommercial-Research
- Code and weights are under the FAIR Noncommercial Research License; the checkpoint is gated on Hugging Face (accept the licence there) and not mirrored.
Our reading of the licences, not legal advice.
Weights
vggt_omega_1b_512.pt 4.6 GB· hosted on Hugging Facefastforks downloads and checks these files on first use (fastforks download vggt-omega-1b-512). 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.
DTU scans 1
9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 %
| Run | dtype | AbsRel ↓ | δ1.25 % ↑ | Δ AbsRel vs fp32 original | scans | ms / scan |
|---|---|---|---|---|---|---|
| Original fp32 | fp32 | 0.0050 | 99.93 | reference | 5 | 2658 |
| Original fp16 | bf16 | 0.0051 | 99.93 | −0.0001 | 5 | 914 |
| Portable fp32 | fp32 | 0.0050 | 99.93 | ±0.0000 | 5 | 2674 |
| Portable | fp16 | 0.0050 | 99.93 | ±0.0000 | 5 | 729 |
| Kernels | fp16 | 0.0051 | 99.93 | −0.0001 | 5 | 465 |
640x512 crops, the 7 most co-visible views as sources; GT = depth rendered from the structured-light points. AbsRel / delta1 over GT pixels, after per-image median scaling unless the model takes the cameras (calibrated MVS: cameras, error in mm; PatchMatchNet also the depth range 425-935 mm, MVSAnywhere its own range from the camera baselines). Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16, portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 30c7a8e9, 2026-11-18.
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| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 and p99 <= 0.25 (absolute error of the output depth).
| Check | Tolerance | Measured |
|---|---|---|
| CPU fp32, max |diff| | ≤ 0.002 | not recorded |
| GPU default dtypes, mean |diff| | ≤ 0.02 | 0.0002 |
| GPU default dtypes, p99 |diff| | ≤ 0.25 | 0.00065 |
Recorded for vggt-omega-1b-512 on RTX 5060 Ti, torch 2.14.1+cu130, 2026-10-31, with these kernels: fastbbvggt_omega
Details
VGGT-Omega (Meta), the successor of VGGT: cameras and depth maps of every view from unposed images in one forward pass. A DINOv3 ViT-L/16 patch embedding feeds 24 frame and 24 inter-frame (global or register-only) attention blocks with q/k norm and 2-D RoPE; a camera head and a DPT dense head decode the tokens. The checkpoint is gated on Hugging Face (accept the FAIR licence there, then put the file on FASTFORKS_WEIGHTS_PATH); there is no mirror.
- Input / output
- Input: several images of one scene (posed and calibrated for multi-view stereo, unposed otherwise). Output: depth maps and, where the model predicts them, cameras and 3-D points.
- Input size
- any H x W >= 1, resized or padded inside the model; tested up to 3000x4000 (5,594 MiB allocated)
- Options
area_side=512>= 16· changes the output
- Outputs
depth,confidence,intrinsics,cam_to_world- CUDA kernels
fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbonesvggt_omegaVGGT-Omega: fused LayerNorm, residual and rotary kernels
- Apple (MPS)
- fp32, fp16
Full reference, generated from the repository
VGGT-Omega (Meta): the successor of VGGT with a DINOv3 backbone, cameras and depth from many images. Non-commercial licence.
Commercial use: no (code: LicenseRef-FAIR-Noncommercial-Research (noncommercial); weights: LicenseRef-FAIR-Noncommercial-Research).
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
vggt-omega-1b-512 | synthetic (336x518) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 2650 906 726 467 589 377 | 5765 / 5912 7495 / 7654 3421 / 3518 3404 / 3494 3082 / 3154 5657 / 5766 |
Details
- Original code: https://github.com/facebookresearch/vggt-omega at
48b23c8 - Code licence: LicenseRef-FAIR-Noncommercial-Research (noncommercial); files:
LICENSE - Weights licence: LicenseRef-FAIR-Noncommercial-Research
- Licence evidence: https://github.com/facebookresearch/vggt-omega/blob/48b23c8/LICENSE and https://huggingface.co/facebook/VGGT-Omega/blob/1041e80fc0e911235d3426b0a3d9a81075111a53/LICENSE.txt
- Training data: Per the repository README: a mixture of public multi-view datasets (uCO3D annotations released); DINOv3 initialisation. The 512 checkpoint is recommended for in-the-wild use, not benchmarking (vggt_omega_1b_416_reproduce.pt is the benchmark reference).
- Licence note: Code and weights are under the FAIR Noncommercial Research License; the checkpoint is gated on Hugging Face (accept the licence there) and not mirrored.
- Options:
area_side= 512 (>= 16; alsoresolution) - Kernel libraries:
fastbb,vggt_omega - Variants:
vggt-omega-1b-512 - Outputs: depth, confidence, intrinsics, cam_to_world; kind:
scale_invariant - Input size: any H x W >= 1, resized or padded inside the model
- Tested up to (portable, fp16, RTX 5060 Ti 16 GB):
vggt-omega-1b-5123000x4000 (5.5 GiB peak allocated, 6.7 reserved) vggt-omega-1b-512336x518, first call: portable + torch.compile 66 s (cold Inductor cache), TensorRT fp16 187 s (ONNX + engine).vggt-omega-1b-512original (fp32) 336x518: float32 (the authors' autocast off)vggt-omega-1b-512original 336x518: original default VGGTOmega(autocast=True): bfloat16 aggregator, float32 headsvggt-omega-1b-512load / first call: original 16.31 s / 1.1 s (336x518), portable 3.31 s / 0.734 s (336x518), kernels 5.24 s / 0.752 s (336x518)- Protocol, median of 9 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
640x512 crops, the 7 most co-visible views as sources; GT = depth rendered from the structured-light points. AbsRel / delta1 over GT pixels, after per-image median scaling unless the model takes the cameras (calibrated MVS: cameras, error in mm; PatchMatchNet also the depth range 425-935 mm, MVSAnywhere its own range from the camera baselines). Original = the original code named under Details, same checkpoint and protocol.
| variant | dataset | original fp32 | original fp16 | portable | portable fp32 | kernels | note |
|---|---|---|---|---|---|---|---|
vggt-omega-1b-512 | DTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 % | 0.0050 / 99.93 | 0.0051 / 99.93 | 0.0050 / 99.93 | 0.0050 / 99.93 | 0.0051 / 99.93 |
Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 30c7a8e9, 2026-11-18; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06.
Parity
CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 and p99 <= 0.25 (absolute error of the output depth).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
vggt-omega-1b-512 | 0.0002 / 0.00065 | RTX 5060 Ti, torch 2.14.1+cu130 | fastbb@7d058322634d:sm_120, vggt_omega@b8fa01afd00c:sm_120 |
Example output
One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).
Input: Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), Scharstein et al., Middlebury College (research dataset, no licence stated).