Models · Multi-view depth

LingBot-Map

LingBot-Map (Robbyant, 2026): streaming 3D reconstruction, cameras and depth for a long image sequence frame by frame with a sliding-window cache. Apache-2.0 code and weights.

experimental Fork of Robbyant/lingbot-map@b9dd1b1 · backends: portable, kernels

Checkpoint lingbot-map

Paper 2026 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 515 ms 480x640, fp16
vs the faster original 1.9× original bf16: 993 ms
Peak memory 4,519 MiB original: 7,141 MiB
Accuracy vs original fp32 ±0.0000 kernels AbsRel, DTU scans 1 (0.0304 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("lingbot-map")  # kernels on a supported GPU, else portable PyTorch
result = model(images)
result.depth  # also: confidence, intrinsics, cam_to_world
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench lingbot-map

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); upstream GCTStream.inference_streaming (demo defaults, SDPA), bfloat16 aggregatorbf16 993–7,141 MiB/7,672 MiB11.52.48
Portable fp32fp32not measured ––––
Portablefp16 7301.4×4,512 MiB/4,770 MiB6.61.08
Kernelsfp16 5151.9×4,519 MiB/4,810 MiB5.70.84
Portable + torch.compileplanned for 0.2.1fails––––
TensorRT fp16fp16 4892.0×3,550 MiBenginenot comparable181 (build)–

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, 2026-11-17Portablefp167125 8,834 MiB processskip480x640
Portablefp328840 12.7 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
Apache-2.0 LICENSE
Weights
Apache-2.0 source
Training data
Per the README: a stage-1 bidirectional (VGGT-style) checkpoint, then streaming training; datasets are not listed

Our reading of the licences, not legal advice.

Weights

lingbot-map.pt 4.6 GB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download lingbot-map).

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 % ↑Δ AbsRel vs fp32 originalscansms / scan
Original fp32fp320.030497.30 reference 5 3284
Original fp16not measured
Portable fp32fp320.030497.30 ±0.0000 5 3197
Portablefp160.030597.28 −0.0001 5 788
Kernelsfp160.030497.33 ±0.0000 5 556

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: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-16; 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 68b2f345, 2026-11-16.

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 0.00042
GPU default dtypes, p99 |diff|≤ 0.25 0.00121

Recorded for lingbot-map on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-14, with these kernels: cudnn_haccfastbb

Details

LingBot-Map (Robbyant, 2026): streaming 3D reconstruction. A VGGT-style aggregator (DINOv2 ViT-L patch embed, 24 frame / global attention blocks, 3D RoPE) runs causally over the views in order with a KV cache (scale frames + a sliding window), then a camera head and a DPT depth head give cameras and depth of every view.

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 (5,593 MiB allocated)
Options
  • long_side=518>= 14· changes the output
  • scale_frames=8>= 1· changes the output
  • window=64>= 1· changes the output
Outputs
depth, confidence, intrinsics, cam_to_world
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

Limitations

  • Views are a sequence (a video): each view only sees the first `scale_frames` views and the views before it (older than `window` frames: their 6 special tokens only). At most 1024 views.
Full reference, generated from the repository

LingBot-Map (Robbyant, 2026): streaming 3D reconstruction, cameras and depth for a long image sequence frame by frame with a sliding-window cache. Apache-2.0 code and weights.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
lingbot-mapsynthetic (480x640)original
portable
kernels
portable + torch.compile
TensorRT fp16
993
730
515
planned for 0.2.1
489
7141 / 7672
4512 / 4770
4519 / 4810
–
6280 / 6338
Details
  • Original code: https://github.com/Robbyant/lingbot-map at b9dd1b1
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0
  • Licence evidence: https://github.com/Robbyant/lingbot-map/blob/b9dd1b1/LICENSE.txt and https://huggingface.co/robbyant/lingbot-map (no licence field)
  • Training data: Per the README: a stage-1 bidirectional (VGGT-style) checkpoint, then streaming training; datasets are not listed
  • Options: long_side = 518 (>= 14), scale_frames = 8 (>= 1), window = 64 (>= 1)
  • Kernel libraries: fastbb, cudnn_hacc
  • Variants: lingbot-map
  • Not in this release (planned for 0.2.1): lingbot-map compile(): Inductor fails to compile a fused Triton kernel (cat / cos / sin)
  • Outputs: depth, confidence, intrinsics, cam_to_world; 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): lingbot-map 3000x4000 (5.5 GiB peak allocated, 6.8 reserved)
  • lingbot-map 480x640, first call: TensorRT fp16 181 s (ONNX + engine).
  • lingbot-map original 480x640: the authors' mixed-precision option (bfloat16); upstream GCTStream.inference_streaming (demo defaults, SDPA), bfloat16 aggregator
  • lingbot-map load / first call: original 131.28 s / 4.645 s (480x640), portable 6.58 s / 1.077 s (480x640), kernels 5.69 s / 0.843 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.

variantdatasetoriginal fp32portableportable fp32kernelsnote
lingbot-mapDTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 %0.0304 / 97.300.0305 / 97.280.0304 / 97.300.0304 / 97.33

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 68b2f345, 2026-11-16; 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 68b2f345, 2026-11-16.

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
lingbot-map0.00042 / 0.00121RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fastbb@7d058322634d:sm_120

Limitations

  • Views are a sequence (a video): each view only sees the first scale_frames views and the views before it (older than window frames: their 6 special tokens only). At most 1024 views.

Example output

No example rendered for this checkpoint yet.