Models · Multi-view depth

ReSplat

ReSplat (ETH Zurich / cvg): 3D Gaussian splats from several posed images, refined by rendering them back; saves a 3DGS .ply. MIT (DL3DV Small), non-commercial (Base).

experimental Fork of cvg/resplat@cae7ddc · backends: portable, kernels

Checkpoint

Paper 2025 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 362 ms 480x640, fp16
vs the faster original 1.4× original bf16: 493 ms
Peak memory 4,981 MiB original: 5,148 MiB
Accuracy vs original fp32 – original not run on the test set

Use it

pip install fastforks

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

from fastforks import multiview

model = multiview.load("resplat-small-dl3dv")  # 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 resplat-small-dl3dv

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, 480x640

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32fp32not measured ––––
Original (own default)the authors' mixed-precision option (bfloat16); cvg/resplat@cae7ddc4cdbd EncoderReSplat, pointops + gsplat extensions, 2 refinement steps, released bfloat16 autocastbf16 493–5,148 MiB/5,376 MiB1.02.68
Portable fp32fp32not measured ––––
Portablefp16 17230.29×4,995 MiB/5,180 MiB0.72.51
Kernelsfp16 3621.4×4,981 MiB/5,148 MiB3.92.84
Portable + torch.compile 15380.32×4,995 MiB/5,138 MiB243 (build)–
TensorRTplanned for 0.2.1fails––––

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-15 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

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-17Portablefp3214422 11.8 GiB processskip480x640
Kernelsfp16not implemented–480x640

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
  • Code and checkpoints are MIT. The Base RealEstate10K checkpoint's monocular branch is initialised from Depth Anything V2 Base (CC-BY-NC-4.0) and is labelled non-commercial; the Small DL3DV checkpoint builds on Depth Anything V2 Small (Apache-2.0).

Our reading of the licences, not legal advice.

Weights

resplat-small-dl3dv-256x448-view8-548993fe.pth 307 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download resplat-small-dl3dv).

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 ↓scansms / scan
Original fp32not measured
Original fp16not measured
Portable fp32fp320.423310.62270.85 5 1466
Portablenot measured
Kernelsnot measured

The original code was not run on this dataset, so the table cannot show the distance to it.

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). Measured on 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| <= 5e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.01 and p99 <= 0.05 (absolute error of the output depth).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.00005 not recorded
cpu_outliers≤ 0.02 not recorded
GPU default dtypes, mean |diff|≤ 0.01 0.00302
GPU default dtypes, p99 |diff|≤ 0.05 0.00738

Recorded for resplat-small-dl3dv on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-14, with these kernels: cudnn_haccfastbbgsraster

Details

ReSplat (Xu et al. 2025): feed-forward 3D Gaussian splatting from posed, calibrated images with recurrent refinement. DepthSplat's multi-view depth network (cost volume + DINOv2) and a kNN point transformer predict latent Gaussians (SH degree 3); refinement steps render the input views, compare them with the images (ResNet-18 features + colour) and update every Gaussian with a second point transformer. Returns every view's depth and the Gaussians in the given world frame (`result.gaussians`, `.save("scene.ply")`).

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 (6,071 MiB allocated)
Options
  • long_side=0>= 0· changes the output
  • refine=2>= 0· changes the output
Outputs
depth, confidence, gaussians
Checkpoint
vit="vits", channels=256, latent=4, fixed_latent=true, gaussians_per_point=1, same_points=false, lowest=8, train_size=[256,448], far=200
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)
  • gsraster3D Gaussian splatting rasteriser
Apple (MPS)
fp32

Limitations

  • No accuracy measurement: a metric model at RealEstate10K / DL3DV scale; the only multi-view eval set here (DTU, millimetre scale) is out of its range, and the original could not run here (its gsplat rasteriser module is not installable), so no tier has a published-protocol score.
  • Portable runs upstream's two CUDA extensions (kNN, gsplat rasteriser) as PyTorch ops: 2.8x (base) / 3.5x (small) slower than the bf16 original at 480x640. Slow: kNN topk ~28 %, compositing cumprod/pow ~30 % of GPU time. Kernels: 2.0x / 1.4x faster.
  • The portable kNN pooling calls torch ops that return non-Tensors: 38 graph breaks (`torch.compile` runs with breaks, no `torch.export`).
Full reference, generated from the repository

ReSplat (ETH Zurich / cvg): 3D Gaussian splats from several posed images, refined by rendering them back; saves a 3DGS .ply. MIT (DL3DV Small), non-commercial (Base).

Commercial use: resplat-small-dl3dv: yes; resplat-base-re10k: no (weights: CC-BY-NC-4.0).

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
resplat-small-dl3dvsynthetic (480x640)original
portable
kernels
portable + torch.compile
TensorRT fp16
493
1723
362
1538
planned for 0.2.1
5148 / 5376
4995 / 5180
4981 / 5148
4995 / 5138
–
resplat-base-re10ksynthetic (480x640)original
portable
kernels
portable + torch.compile
TensorRT fp16
2521
7086
1280
fails: timeout after 900 s (audit 2026-11-16, RTX 5060 Ti: compile
fails: TypeError: _Base.forward() missing 1 required positional arg
10139 / 12684
6951 / 8736
6944 / 8706
–
–
Details
  • Original code: https://github.com/cvg/resplat at cae7ddc
  • Code licence: MIT (permissive); files: LICENSE
  • Weights licence: MIT
    • resplat-base-re10k: CC-BY-NC-4.0
  • Licence evidence: https://github.com/cvg/resplat/blob/cae7ddc/LICENSE and https://github.com/cvg/resplat/blob/cae7ddc/MODEL_ZOO.md and https://huggingface.co/haofeixu/resplat/tree/c69ad72a49b8358ac46599cf74468090425ed9c7
  • Training data: DL3DV-10K (multi-view real-world scenes for 3D vision) (DL3DV-10K (non-commercial research licence), from DepthSplat / Depth Anything V2 Small depth initialisations.)
    • resplat-base-re10k: RealEstate10K (camera trajectories of YouTube real-estate videos) (RealEstate10K (YouTube videos, camera poses from ORB-SLAM; research use per its release), from a DepthSplat / Depth Anything V2 Base (CC-BY-NC-4.0) depth initialisation.)
  • Licence note: Code and checkpoints are MIT. The Base RealEstate10K checkpoint's monocular branch is initialised from Depth Anything V2 Base (CC-BY-NC-4.0) and is labelled non-commercial; the Small DL3DV checkpoint builds on Depth Anything V2 Small (Apache-2.0).
  • Options: long_side = 0 (>= 0), refine = 2 (>= 0)
  • Kernel libraries: fastbb, cudnn_hacc, gsraster
  • Variants: resplat-small-dl3dv, resplat-base-re10k
  • Not in this release (planned for 0.2.1): resplat-small-dl3dv ONNX / TensorRT export: torch.export cannot capture data-dependent Gaussian renderer shapes
  • 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): resplat-small-dl3dv 3000x4000 (5.9 GiB peak allocated, 7.0 reserved); resplat-base-re10k 3000x4000 (7.8 GiB peak allocated, 11.2 reserved)
  • resplat-small-dl3dv 480x640, first call: portable + torch.compile 243 s (cold Inductor cache).
  • resplat-small-dl3dv original 480x640: the authors' mixed-precision option (bfloat16); cvg/resplat@cae7ddc4cdbd EncoderReSplat, pointops + gsplat extensions, 2 refinement steps, released bfloat16 autocast
  • resplat-base-re10k original 480x640: the authors' mixed-precision option (bfloat16); cvg/resplat@cae7ddc4cdbd EncoderReSplat, pointops + gsplat extensions, 2 refinement steps, released bfloat16 autocast
  • resplat-small-dl3dv load / first call: original 16.8 s / 10.688 s (480x640), portable 0.72 s / 2.51 s (480x640), kernels 3.88 s / 2.839 s (480x640)
  • resplat-base-re10k load / first call: original 2.26 s / 2.743 s (480x640), portable 0.9 s / 7.756 s (480x640), kernels 3.3 s / 1.98 s (480x640)
  • 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.

variantdatasetportable fp32note
resplat-small-dl3dvDTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 % / MAE mm0.4233 / 10.62 / 270.85
resplat-base-re10kDTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 % / MAE mm0.1396 / 98.14 / 92.13

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 30c7a8e9, 2026-11-18.

Parity

CPU, float32: max |diff| <= 5e-05 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.01 and p99 <= 0.05 (absolute error of the output depth).

variantkernels error vs original, mean / p99GPU, torchkernels used
resplat-small-dl3dv0.00302 / 0.00738RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fastbb@7d058322634d:sm_120, gsraster@be9d99ba7071:sm_120
resplat-base-re10k0.00424 / 0.01598RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fastbb@7d058322634d:sm_120, gsraster@be9d99ba7071:sm_120

Limitations

  • No accuracy measurement: a metric model at RealEstate10K / DL3DV scale; the only multi-view eval set here (DTU, millimetre scale) is out of its range, and the original could not run here (its gsplat rasteriser module is not installable), so no tier has a published-protocol score.
  • Portable runs upstream's two CUDA extensions (kNN, gsplat rasteriser) as PyTorch ops: 2.8x (base) / 3.5x (small) slower than the bf16 original at 480x640. Slow: kNN topk ~28 %, compositing cumprod/pow ~30 % of GPU time. Kernels: 2.0x / 1.4x faster.
  • The portable kNN pooling calls torch ops that return non-Tensors: 38 graph breaks (torch.compile runs with breaks, no torch.export).

Example output

No example rendered for this checkpoint yet.