Models · Multi-view depth

DepthSplat

DepthSplat (ETH Zurich / cvg): 3D Gaussian splats and depth from several posed images in one pass; saves a 3DGS .ply. MIT (Small checkpoint).

experimental Fork of cvg/depthsplat@2dad25a · backends: portable, kernels

Checkpoint depthsplat-small-re10k

Paper 2024 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 112 ms 336x518, fp16
vs the faster original 1.1× original fp16: 126 ms
Peak memory 5,434 MiB original: 5,694 MiB
Accuracy vs original fp32 +0.0018 kernels AbsRel, DTU scans 1 (0.0992 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("depthsplat-small-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 depthsplat-small-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); cvg/depthsplat@2dad25a7b9ba EncoderDepthSplat, long side 256fp16 126–5,694 MiB/5,892 MiB0.50.35
Portable fp32fp32not measured ––––
Portablefp16 1191.1×5,432 MiB/5,606 MiB0.20.97
Kernelsfp16 1121.1×5,434 MiB/5,606 MiB1.10.80
Portable + torch.compile 82.21.5×5,432 MiB/5,526 MiB80 (build)–
TensorRT fp16fp16 86.41.5×2,952 MiBenginenot comparable224 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-12 to 2026-11-15. 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–336x518
Portablefp16757 8,900 MiB processmatches336x518
Portablefp32670 8,880 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 Small RealEstate10K checkpoint is shipped: MIT code and weights, its monocular branch initialised from Depth Anything V2 Small (Apache-2.0). The Base / Large checkpoints build on Depth Anything V2 Base / Large (CC-BY-NC-4.0) and are not offered.

Our reading of the licences, not legal advice.

Weights

depthsplat-gs-small-re10k-256x256-view2-cfeab6b1.pth 154 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download depthsplat-small-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 fp32fp320.099298.7266.52 reference 5 238
Original fp16fp160.097498.6965.37 +0.0018 5 173
Portable fp32fp320.099298.7266.52 ±0.0000 5 225
Portablefp160.097498.7065.36 +0.0018 5 164
Kernelsfp160.097498.6965.34 +0.0018 5 155

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, 2026-11-12; portable fp32, portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 62af007c, 2026-11-12.

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| <= 1.6e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0035 and p99 <= 0.017 (absolute error of the output depth).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.000016 not recorded
GPU default dtypes, mean |diff|≤ 0.0035 0.00197
GPU default dtypes, p99 |diff|≤ 0.017 0.00899

Recorded for depthsplat-small-re10k on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12, with these kernels: cudnn_haccfastbb

Details

DepthSplat (Xu et al. 2025): feed-forward 3D Gaussian splatting and depth from posed, calibrated images. A multi-view cost volume (UniMatch transformer, 2D U-Net) fused with a monocular DINOv2 ViT-S predicts every view's depth; a DPT feature upsampler and a Gaussian head predict one Gaussian per pixel (opacity, scale, rotation, degree-2 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 (7,581 MiB allocated)
Options
  • long_side=256>= 16· changes the output
Outputs
depth, confidence, gaussians
CUDA kernels
  • fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbones
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

DepthSplat (ETH Zurich / cvg): 3D Gaussian splats and depth from several posed images in one pass; saves a 3DGS .ply. MIT (Small checkpoint).

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
depthsplat-small-re10ksynthetic (336x518)original
portable
kernels
portable + torch.compile
TensorRT fp16
126
119
112
82.2
86.4
5694 / 5892
5432 / 5606
5434 / 5606
5432 / 5526
3148 / 3194
Details
  • Original code: https://github.com/cvg/depthsplat at 2dad25a
  • Code licence: MIT (permissive); files: LICENSE
  • Weights licence: MIT
  • Licence evidence: https://github.com/cvg/depthsplat/blob/2dad25a7b9ba7c5537aba08463ef88d8511a06be/LICENSE and https://github.com/cvg/depthsplat/blob/2dad25a7b9ba7c5537aba08463ef88d8511a06be/MODEL_ZOO.md and https://huggingface.co/haofeixu/depthsplat/tree/ed5116c11f7932fff3714989e2b65716730f6f87
  • Training data: RealEstate10K (camera trajectories of YouTube real-estate videos) (RealEstate10K (YouTube videos, camera poses from ORB-SLAM; research use per its release), from a UniMatch depth / Depth Anything V2 Small initialisation.)
  • Licence note: Only the Small RealEstate10K checkpoint is shipped: MIT code and weights, its monocular branch initialised from Depth Anything V2 Small (Apache-2.0). The Base / Large checkpoints build on Depth Anything V2 Base / Large (CC-BY-NC-4.0) and are not offered.
  • Options: long_side = 256 (>= 16)
  • Kernel libraries: fastbb, cudnn_hacc
  • Variants: depthsplat-small-re10k
  • 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): depthsplat-small-re10k 3000x4000 (7.4 GiB peak allocated, 8.6 reserved)
  • depthsplat-small-re10k 336x518, first call: portable + torch.compile 80 s (cold Inductor cache), TensorRT fp16 224 s (ONNX + engine).
  • depthsplat-small-re10k original 336x518: the authors' mixed-precision option (float16); cvg/depthsplat@2dad25a7b9ba EncoderDepthSplat, long side 256
  • depthsplat-small-re10k load / first call: original 1.83 s / 2.968 s (336x518), portable 0.22 s / 0.968 s (336x518), kernels 1.47 s / 0.71 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
depthsplat-small-re10kDTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 % / MAE mm0.0992 / 98.72 / 66.520.0974 / 98.69 / 65.370.0974 / 98.70 / 65.360.0992 / 98.72 / 66.520.0974 / 98.69 / 65.34

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 62af007c, 2026-11-12.

Parity

CPU, float32: max |diff| <= 1.6e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0035 and p99 <= 0.017 (absolute error of the output depth).

variantkernels error vs original, mean / p99GPU, torchkernels used
depthsplat-small-re10k0.00197 / 0.00899RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fastbb@7d058322634d:sm_120

Example output

No example rendered for this checkpoint yet.