VGGT
VGGT (Meta): cameras, depth and 3D points for many images in one pass. Non-commercial weights.
stable Fork of facebookresearch/vggt@a288dd0 · backends: portable, kernels
- Paper VGGT: Visual Geometry Grounded Transformer CVPR 2025 · first released 14 Mar 2025
Checkpoint vggt-1b
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 multiview
model = multiview.load("vggt-1b") # kernels on a supported GPU, else portable PyTorch
result = model(images)
result.depth # also: confidence, world_points, intrinsics, cam_to_world, points_confidence
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-1b 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 | 2502 | 0.39× | 6,958 MiB/7,358 MiB | 14.2 | 2.50 |
| Original fp16original default (README / demo): torch.autocast(bfloat16) | fp16 | 965 | – | 8,688 MiB/9,158 MiB | 19.4 | 1.29 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 651 | 1.5× | 3,643 MiB/3,874 MiB | 6.8 | 1.02 |
| Kernels | fp16 | 460 | 2.1× | 3,650 MiB/4,194 MiB | 7.0 | 0.75 |
| Portable + torch.compile | 557 | 1.7× | 3,611 MiB/3,856 MiB | 74 (build) | – | |
| TensorRT fp16 | fp16 | 424 | 2.3× | 3,642 MiBenginenot comparable | 199 (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, 2026-11-14 | Original | fp32 | 7987 | 10.1 GiB process | matches | 336x518 |
| Portable | fp16 | 6638 | 8,423 MiB process | matches | 336x518 | |
| Portable | fp32 | 7209 | 10.4 GiB process | matches | 336x518 | |
| Kernels | fp32 | 7016 | 10.4 GiB process | matches | 336x518 | |
| NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 663 | 3,619 MiB/3,834 MiB | – | 336x518 |
| Kernels | fp16 | 660 | 3,640 MiB/4,134 MiB | – | 336x518 | |
| NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 196 | 3,619 MiB/3,834 MiB | – | 336x518 |
| Kernels | fp16 | 189 | 3,640 MiB/4,134 MiB | – | 336x518 | |
| NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07 | Portable | fp16 | 517 | 3,619 MiB/3,834 MiB | – | 336x518 |
| Kernels | fp16 | 504 | 3,630 MiB/4,174 MiB | – | 336x518 | |
| NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 190 | 3,619 MiB/3,834 MiB | – | 336x518 |
| Kernels | fp16 | 184 | 3,634 MiB/4,134 MiB | – | 336x518 | |
| NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08 | Portable | fp16 | 104 | 3,643 MiB/3,874 MiB | – | 336x518 |
| Kernels | fp16 | 106 | 3,654 MiB/4,194 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-VGGT LICENSE
- Weights
- CC-BY-NC-4.0 source
- Training data
- Per the paper (arXiv 2503.11651): Co3Dv2, BlendMVS, DL3DV, MegaDepth, Kubric, WildRGB, ScanNet, HyperSim, Mapillary, Habitat, Replica, MVS-Synth, PointOdyssey, Virtual KITTI, Aria Synthetic Environments, Aria Digital Twin; DINOv2 initialisation
- weights: CC-BY-NC-4.0
- The VGGT code licence allows commercial use, but the VGGT-1B weights are CC-BY-NC-4.0; Meta's separately licensed commercial checkpoint is not packaged.
Our reading of the licences, not legal advice.
Weights
model.pt 5.0 GB· hosted on Hugging Facefastforks downloads and checks these files on first use (fastforks download vggt-1b). 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 | 0.0026 | 99.94 | reference | 5 | 3422 |
| Original fp16 | bf16 | 0.0027 | 99.94 | −0.0001 | 5 | 1253 |
| Portable fp32 | fp32 | 0.0026 | 99.94 | ±0.0000 | 5 | 3400 |
| Portable | fp16 | 0.0026 | 99.94 | ±0.0000 | 5 | 856 |
| Kernels | fp16 | 0.0026 | 99.94 | ±0.0000 | 5 | 604 |
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, 2026-11-17.
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 | 0.00044 |
| GPU default dtypes, p99 |diff| | ≤ 0.25 | 0.00183 |
Recorded for vggt-1b on RTX 5060 Ti, torch 2.14.1+cu130, 2026-10-31, with these kernels: cudnn_haccfastbb
Details
VGGT, Visual Geometry Grounded Transformer (Wang et al. 2025): cameras, depth maps and point maps of every view from unposed images in one forward pass. A DINOv2 ViT-L patch embedding feeds 24 alternating frame / global attention blocks; a camera head and two DPT heads decode the tokens.
- 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 (8,219 MiB allocated)
- Options
long_side=518>= 14· changes the output
- Outputs
depth,confidence,world_points,intrinsics,cam_to_world,points_confidence(on request)- CUDA kernels
fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbonescudnn_haccconvolutions with fp16 accumulation (cuDNN)
- Apple (MPS)
vggt_metalMetal: VGGT rotary embeddings. fp32, fp16
Full reference, generated from the repository
VGGT (Meta): cameras, depth and 3D points for many images in one pass. Non-commercial weights.
Commercial use: no (weights: CC-BY-NC-4.0).
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
vggt-1b | synthetic (336x518) | original (fp32) original portable kernels portable + torch.compile TensorRT fp16 | 2502 965 651 460 557 424 | 6958 / 7358 8688 / 9158 3643 / 3874 3650 / 4194 3611 / 3856 6421 / 6472 |
Details
- Original code: https://github.com/facebookresearch/vggt at
a288dd0 - Code licence: LicenseRef-VGGT (permissive); files:
LICENSE - Weights licence: CC-BY-NC-4.0
- Licence evidence: https://github.com/facebookresearch/vggt/blob/a288dd0/LICENSE.txt and https://huggingface.co/facebook/VGGT-1B (license: cc-by-nc-4.0)
- Training data: Per the paper (arXiv 2503.11651): Co3Dv2, BlendMVS, DL3DV, MegaDepth, Kubric, WildRGB, ScanNet, HyperSim, Mapillary, Habitat, Replica, MVS-Synth, PointOdyssey, Virtual KITTI, Aria Synthetic Environments, Aria Digital Twin; DINOv2 initialisation
- Licence note: The VGGT code licence allows commercial use, but the VGGT-1B weights are CC-BY-NC-4.0; Meta's separately licensed commercial checkpoint is not packaged.
- Options:
long_side= 518 (>= 14) - Kernel libraries:
fastbb,cudnn_hacc - Variants:
vggt-1b - Outputs: depth, confidence, world_points, intrinsics, cam_to_world, points_confidence (in
extras); 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-1b3000x4000 (8.0 GiB peak allocated, 10.1 reserved) vggt-1b336x518, first call: portable + torch.compile 74 s (cold Inductor cache), TensorRT fp16 199 s (ONNX + engine).vggt-1boriginal (fp32) 336x518: float32 (the authors' autocast off)vggt-1boriginal 336x518: original default (README / demo): torch.autocast(bfloat16)vggt-1bload / first call: original 19.37 s / 1.289 s (336x518), portable 6.78 s / 1.025 s (336x518), kernels 7.03 s / 0.751 s (336x518)- Protocol, 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
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-1b | DTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 % | 0.0026 / 99.94 | 0.0027 / 99.94 | 0.0026 / 99.94 | 0.0026 / 99.94 | 0.0026 / 99.94 |
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, 2026-11-17; 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-1b | 0.00044 / 0.00183 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_hacc@cudnn92400, fastbb@7d058322634d: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).