Models · Multi-view depth

MVSAnywhere

MVSAnywhere (Niantic): depth of a reference view from several posed images, for any scene type. Non-commercial licence.

stable Fork of nianticlabs/mvsanywhere@5bd49bb · backends: portable, kernels

Checkpoint mvsanywhere

Paper 2025 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 35.2 ms 448x672, fp16
vs the faster original 6.2× original fp16: 219 ms
Peak memory 702 MiB original: 1,504 MiB
Accuracy vs original fp32 ±0.0000 kernels AbsRel, DTU scans 1 (0.0077 original)

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("mvsanywhere")  # kernels on a supported GPU, else portable PyTorch
result = model(images)
result.depth  # also: mask, sky, depth_range
result.backend  # what ran: mode, device, dtype, kernels

The weights are non-commercial: fastforks downloads them only with FASTFORKS_NONCOMMERCIAL=1 set.

fastforks bench mvsanywhere

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, 448x672

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32 2730.80×1,441 MiB/1,500 MiB15.21.05
Original fp16torch.autocast(float16) around the callfp16 219–1,504 MiB/1,576 MiB3.21.04
Portable fp32fp32not measured ––––
Portablefp16 93.22.3×931 MiB/1,226 MiB1.90.22
Kernelsfp16 35.26.2×702 MiB/760 MiB1.80.48
Portable + torch.compile 70.83.1×818 MiB/860 MiB153 (build)–
TensorRTplanned for 0.2.1fails––––

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-03 to 2026-11-12. 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

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-14Originalfp32cannot run here–448x672
Portablefp161064 3,801 MiB processmatches448x672
Portablefp321051 3,534 MiB processmatches448x672
Kernelsfp16not implemented–448x672
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16126 847 MiB/948 MiB–448x672
Kernelsfp1673.2 744 MiB/780 MiB–448x672
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1659.2 847 MiB/968 MiB–448x672
Kernelsfp1629.3 996 MiB/1,060 MiB–448x672
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp16111 847 MiB/948 MiB–448x672
Kernelsfp1671.9 786 MiB/840 MiB–448x672
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1657.8 847 MiB/948 MiB–448x672
Kernelsfp1656.9 1,004 MiB/1,140 MiB–448x672
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp1623.6 871 MiB/988 MiB–448x672
Kernelsfp1612.5 1,134 MiB/1,180 MiB–448x672

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-noncommercial AND CC-BY-NC-SA-4.0 LICENSELICENSE-antialiased-cnns.md
Weights
LicenseRef-noncommercial AND CC-BY-NC-SA-4.0 (assumed from code)(not stated by the authors; assumed to follow the code) source
Training data
Synthetic data per the paper (arXiv 2411.14390): Hypersim, TartanAir, BlendedMVG, MatrixCity, VKITTI2, Dynamic Replica, MVS-Synth, SAIL-VOS 3D; DINOv2 / Depth Anything V2 initialisation
  • code: LicenseRef-noncommercial AND CC-BY-NC-SA-4.0 (noncommercial)
  • weights: LicenseRef-noncommercial AND CC-BY-NC-SA-4.0 (assumed from code)
  • weights licence not stated by the original authors; assumed to be the code licence
  • Niantic's non-commercial licence covers the code and the checkpoint; the BlurPool stem comes from Adobe's antialiased-cnns (CC-BY-NC-SA-4.0).

Our reading of the licences, not legal advice.

Weights

mvsanywhere_hero.ckpt 2.4 GB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download mvsanywhere). 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 %

RundtypeAbsRel ↓δ1.25 % ↑MAE mm ↓Δ AbsRel vs fp32 originalscansms / scan
Original fp32fp320.007799.825.57 reference 5 315
Original fp16fp160.011099.847.85 −0.0033 5 233
Portable fp32fp320.007799.825.57 ±0.0000 5 236
Portablefp160.007899.825.60 −0.0001 5 115
Kernelsfp160.007799.825.59 ±0.0000 5 44.2

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: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-06; portable fp32, portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 555d7b79, 2026-11-06.

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| <= 0.1 % and p99 <= 0.5 % of the reference's median |value| (absolute error of the output depth).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.002 (relative)not recorded
GPU default dtypes, mean |diff|≤ 0.001 (relative)0.00005 (relative)
GPU default dtypes, p99 |diff|≤ 0.005 (relative)0.00023 (relative)

Recorded for mvsanywhere on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12, with these kernels: cudnn_hacccutlass_convdpt_fusedfastbbigevmvsanywhere

Details

MVSAnywhere (Izquierdo et al. 2025): zero-shot multi-view stereo, the depth of a reference view from posed, calibrated source views. A view-agnostic MLP cost volume (warped matching features + camera metadata, softmax over the views) is encoded by a DINOv2 ViT-B fused with a second DINOv2's image features and decoded by a DPT head; the depth range comes from the camera baselines and is refined once.

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 (3,110 MiB allocated)
Options
  • long_side=640>= 64· changes the output
  • refinement_steps=1>= 0· changes the output
Outputs
depth, mask, sky(on request), depth_range(on request)
CUDA kernels
  • fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbones
  • igevIGEV stereo: convolution epilogues, correlation lookups, GRU gates
  • cutlass_convCUTLASS convolutions with the recurrent update (GRU) fused in
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • mvsanywhereMVSAnywhere: fused cost-volume MLP
  • dpt_fusedfused fp16 matrix multiplies and convolutions for DINOv2 backbones and DPT depth heads
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

MVSAnywhere (Niantic): depth of a reference view from several posed images, for any scene type. Non-commercial licence.

Commercial use: no (code: LicenseRef-noncommercial AND CC-BY-NC-SA-4.0 (noncommercial); weights: LicenseRef-noncommercial AND CC-BY-NC-SA-4.0 (assumed from code)).

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
mvsanywheresynthetic (448x672)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
273
219
93.2
35.2
70.8
planned for 0.2.1
1441 / 1500
1504 / 1576
931 / 1226
702 / 760
818 / 860
–
Details
  • Original code: https://github.com/nianticlabs/mvsanywhere at 5bd49bb
  • Code licence: LicenseRef-noncommercial AND CC-BY-NC-SA-4.0 (noncommercial); files: LICENSE, LICENSE-antialiased-cnns.md
  • Weights licence: LicenseRef-noncommercial AND CC-BY-NC-SA-4.0 (assumed from code)
  • Licence evidence: https://github.com/nianticlabs/mvsanywhere/blob/5bd49bb/LICENSE and README.md (the checkpoint is published under the repository's licence)
  • Training data: Synthetic data per the paper (arXiv 2411.14390): Hypersim, TartanAir, BlendedMVG, MatrixCity, VKITTI2, Dynamic Replica, MVS-Synth, SAIL-VOS 3D; DINOv2 / Depth Anything V2 initialisation
  • Licence note: Niantic's non-commercial licence covers the code and the checkpoint; the BlurPool stem comes from Adobe's antialiased-cnns (CC-BY-NC-SA-4.0).
  • Options: long_side = 640 (>= 64), refinement_steps = 1 (>= 0)
  • Kernel libraries: fastbb, igev, cutlass_conv, cudnn_hacc, mvsanywhere, dpt_fused
  • Variants: mvsanywhere
  • Not in this release (planned for 0.2.1): mvsanywhere TensorRT export: TensorRT produces all-NaN depth and sky maps in both FP16 and FP32
  • Outputs: depth, mask, sky (in extras), depth_range (in extras); kind: metric
  • Input size: any H x W >= 1, resized or padded inside the model
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): mvsanywhere 3000x4000 (3.0 GiB peak allocated, 4.2 reserved)
  • mvsanywhere 448x672, first call: portable + torch.compile 153 s (cold Inductor cache).
  • mvsanywhere original 448x672: torch.autocast(float16) around the call
  • mvsanywhere load / first call: original 3.21 s / 1.037 s (448x672), portable 1.88 s / 0.221 s (448x672), kernels 1.76 s / 0.48 s (448x672)
  • 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.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
mvsanywhereDTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 % / MAE mm0.0077 / 99.82 / 5.570.0110 / 99.84 / 7.850.0078 / 99.82 / 5.600.0077 / 99.82 / 5.570.0077 / 99.82 / 5.59

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 555d7b79, 2026-11-06.

Parity

CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.1 % and p99 <= 0.5 % of the reference's median |value| (absolute error of the output depth).

variantkernels error vs original, mean / p99GPU, torchkernels used
mvsanywhere5e-05 / 0.00023RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, cutlass_conv@988138e042cb:sm_120, dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120, igev@f075827d68c6:sm_120, mvsanywhere@7fdcb3aafdd8: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).

View 1
View 2
Output
Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), view 1
View 1 · 100%

Input: Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), Scharstein et al., Middlebury College (research dataset, no licence stated).