Models · Multi-view depth

MVSplat

MVSplat (Chen et al.): 3D Gaussian splats and depth from several posed, calibrated images in one pass; saves a 3DGS .ply. MIT.

experimental Fork of donydchen/mvsplat@01f9a28 · backends: portable, kernels

Checkpoint mvsplat-re10k

Paper 2024 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 172 ms 336x518, fp16
vs the faster original 1.6× original fp16: 279 ms
Peak memory 1,239 MiB original: 2,889 MiB
Accuracy vs original fp32 ±0.0000 kernels AbsRel, DTU scans 1 (0.0438 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install fastforks-core

from fastforks import multiview

model = multiview.load("mvsplat-re10k")  # kernels on a supported GPU, else portable PyTorch
result = model(images)
result.depth  # also: confidence, gaussians
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench mvsplat-re10k

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

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32not measured ––––
Original fp16the authors' mixed-precision option (float16); donydchen/mvsplat@01f9a28edb5e EncoderCostVolume, long side 256fp16 279–2,889 MiB/4,160 MiB0.20.56
Portable fp32fp32not measured ––––
Portablemeasured on 45b0e4f0 with the cudnn_hacc per-call workspace change of 600d8c2cfp16 1811.5×1,239 MiB/1,532 MiB0.11.02
Kernelsmeasured on 45b0e4f0 with the cudnn_hacc per-call workspace change of 600d8c2cfp16 1721.6×1,239 MiB/1,532 MiB1.30.69
Portable + torch.compile 1152.4×2,762 MiB/2,810 MiB102 (build)–
TensorRTplanned for 0.2.1fails––––

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-13. 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, 2026-11-13Originalfp32cannot run here–336x518
Portablefp161112 4,537 MiB processmatches336x518
Portablefp321077 4,523 MiB processmatches336x518
Kernelsfp16not implemented–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

Commercial use allowed

Code
MIT LICENSE
Weights
MIT source
Training data
  • Only the RealEstate10K checkpoint is shipped (MIT code and weights); the ACID checkpoint has the same architecture.

Our reading of the licences, not legal advice.

Weights

re10k.ckpt 48.0 MB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download mvsplat-re10k).

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 fp32fp320.043897.3131.17 reference 5 481
Original fp16fp160.043997.3031.21 −0.0001 5 383
Portable fp32fp320.043897.3131.17 ±0.0000 5 414
Portablefp160.043897.3031.20 ±0.0000 5 254
Kernelsfp160.043897.3031.20 ±0.0000 5 243

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, 2026-11-13; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 30c7a8e9, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 28db82cb, 2026-11-13.

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).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.002 not recorded
GPU default dtypes, mean |diff|≤ 0.02 0.00157
GPU default dtypes, p99 |diff|≤ 0.25 0.00705

Recorded for mvsplat-re10k on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13, with these kernels: cudnn_hacc

Details

MVSplat (Chen et al. 2024): feed-forward 3D Gaussian splatting and depth from posed, calibrated images. A UniMatch CNN and multi-view swin transformer feed a plane-sweep cost volume (all other views), refined by 2D U-Nets with cross-view attention into every view's depth; a Gaussian head predicts one Gaussian per pixel (opacity, scale, rotation, degree-4 SH) in the given world frame (`result.gaussians`, `.save("scene.ply")`). Forward pass only: no renderer.

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 (3,276 MiB allocated)
Options
  • long_side=256>= 16· changes the output
Outputs
depth, confidence, gaussians
CUDA kernels
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

MVSplat (Chen et al.): 3D Gaussian splats and depth from several posed, calibrated images in one pass; saves a 3DGS .ply. MIT.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
mvsplat-re10ksynthetic (336x518)original
portable
kernels
portable + torch.compile
TensorRT fp16
279
181
172
115
planned for 0.2.1
2889 / 4160
1239 / 1532
1239 / 1532
2762 / 2810
–
Details
  • Original code: https://github.com/donydchen/mvsplat at 01f9a28
  • Code licence: MIT (permissive); files: LICENSE
  • Weights licence: MIT
  • Licence evidence: https://github.com/donydchen/mvsplat/blob/01f9a28edb5eb68416e7e63b01f8d90c3bdfbf01/LICENSE and https://github.com/donydchen/mvsplat/blob/01f9a28edb5eb68416e7e63b01f8d90c3bdfbf01/README.md and https://drive.google.com/drive/folders/14_E_5R6ojOWnLSrSVLVEMHnTiKsfddjU
  • Training data: RealEstate10K (camera trajectories of YouTube real-estate videos) (RealEstate10K (YouTube videos, camera poses from ORB-SLAM; research use per its release), the backbone initialised from UniMatch (GMDepth, ScanNet).)
  • Licence note: Only the RealEstate10K checkpoint is shipped (MIT code and weights); the ACID checkpoint has the same architecture.
  • Options: long_side = 256 (>= 16)
  • Kernel libraries: cudnn_hacc
  • Variants: mvsplat-re10k
  • Not in this release (planned for 0.2.1): mvsplat-re10k TensorRT export: the TensorRT engine has not yet been validated against the portable model
  • Outputs: depth, confidence, gaussians; 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): mvsplat-re10k 3000x4000 (3.2 GiB peak allocated, 4.4 reserved)
  • mvsplat-re10k 336x518, first call: portable + torch.compile 102 s (cold Inductor cache).
  • mvsplat-re10k original 336x518: the authors' mixed-precision option (float16); donydchen/mvsplat@01f9a28edb5e EncoderCostVolume, long side 256
  • mvsplat-re10k portable 336x518, kernels 336x518: measured on 45b0e4f0 with the cudnn_hacc per-call workspace change of 600d8c2c
  • mvsplat-re10k load / first call: original 1.37 s / 2.817 s (336x518), portable 0.15 s / 1.019 s (336x518), kernels 1.27 s / 0.693 s (336x518)
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

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
mvsplat-re10kDTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 % / MAE mm0.0438 / 97.31 / 31.170.0439 / 97.30 / 31.210.0438 / 97.30 / 31.200.0438 / 97.31 / 31.170.0438 / 97.30 / 31.20

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 28db82cb, 2026-11-13; 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, fastforks 28db82cb, 2026-11-13.

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).

variantkernels error vs original, mean / p99GPU, torchkernels used
mvsplat-re10k0.00157 / 0.00705RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400

Example output

No example rendered for this checkpoint yet.