Models · Monocular depth

Depth Anything 3

ByteDance's Depth Anything 3: depth from one image, with relative, metric and multi-view versions. The licence differs by variant, see the table.

stable Fork of ByteDance-Seed/Depth-Anything-3@3d835ec · backends: portable, kernels

Checkpoint

Paper 2025 Added in 0.1.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 14.9 ms 994x1482, fp16
vs the faster original 5.7× original fp16: 85.2 ms
Peak memory 805 MiB original: 2,997 MiB
Accuracy vs original fp32 ±0.0000 kernels AbsRel, NYUv2 test (0.0402 original)

Use it

pip install fastforks

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

from fastforks import depth

model = depth.load("da3mono-large")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.depth  # also: confidence, intrinsics, mask, sky
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench da3mono-large

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.

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32float32 (the authors' autocast off)fp32 1270.67×2,419 MiB/2,450 MiB1.80.13
Original fp16original default (api.py): torch.autocast(bfloat16)fp16 85.2–2,997 MiB/3,052 MiB3.00.28
Portable fp32fp32not measured ––––
Portablefp16 24.13.5×814 MiB/868 MiB2.40.03
Kernelsfp16 14.95.7×805 MiB/868 MiB2.20.02
Portable + torch.compile 21.24.0×792 MiB/828 MiB25 (build)–
TensorRT fp16fp16 14.45.9×754 MiBenginenot comparable52 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-03 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, 2026-11-13Originalfp32271 3,393 MiB processmatches994x1482
Portablefp16280 2,499 MiB processmatches994x1482
Portablefp32293 2,517 MiB processmatches994x1482
Kernelsfp16not implemented–994x1482
NVIDIA L4 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1622.6 787 MiB/828 MiB–994x1482
Kernelsfp1619.2 783 MiB/828 MiB–994x1482
NVIDIA L40S CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp167.80 787 MiB/828 MiB–994x1482
Kernelsfp167.66 783 MiB/828 MiB–994x1482
NVIDIA A10 CUDA, torch 2.14.1+cu130, 2026-11-07Portablefp1622.0 787 MiB/828 MiB–994x1482
Kernelsfp1618.5 783 MiB/828 MiB–994x1482
NVIDIA A100-SXM4-40GB CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp1615.5 787 MiB/828 MiB–994x1482
Kernelsfp1612.8 783 MiB/828 MiB–994x1482
NVIDIA RTX PRO 6000 CUDA, torch 2.14.1+cu130, 2026-11-08Portablefp165.94 811 MiB/868 MiB–994x1482
Kernelsfp165.81 807 MiB/848 MiB–994x1482

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
Public academic datasets only (model cards); teacher-student training on synthetic and real multi-view data per the paper
  • The Mono and Metric weights are Apache-2.0 but their training data terms are unchecked (a note only); the other variants' weights are CC-BY-NC-4.0 (`no`).

Our reading of the licences, not legal advice.

Weights

DA3MONO-LARGE/model.safetensors 1.3 GB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download da3mono-large).

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.

NYUv2 test

AbsRel / delta1

RundtypeAbsRel ↓δ1 ↑Δ AbsRel vs fp32 originalimagesms / image
Original fp32fp320.04020.9821 reference 654 94.2
Original fp16bf160.04020.9820 ±0.0000 654 43.0
Portable fp32fp320.04020.9821 ±0.0000 654 92.0
Portablefp160.04020.9821 ±0.0000 654 26.9
Kernelsfp160.04020.9821 ±0.0000 654 16.2

AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. 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: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6d3bab64, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-06.

iBims-1

AbsRel / delta1

RundtypeAbsRel ↓δ1 ↑Δ AbsRel vs fp32 originalimagesms / image
Original fp32fp320.03110.9872 reference 100 93.4
Original fp16bf160.03110.9872 ±0.0000 100 42.5
Portable fp32fp320.03110.9872 ±0.0000 100 91.6
Portablefp160.03110.9872 ±0.0000 100 26.7
Kernelsfp160.03110.9873 ±0.0000 100 15.4

AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. 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: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6d3bab64, 2026-11-17; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-06.

ETH3D two-view train left

AbsRel / delta1training data not stated: possibly seen

RundtypeAbsRel ↓δ1 ↑Δ AbsRel vs fp32 originalimagesms / image
Original fp32fp320.06310.9567 reference 27 71.6
Original fp16bf160.06290.9583 +0.0002 27 38.0
Portable fp32not measured
Portablefp160.06300.9568 +0.0001 27 22.2
Kernelsfp160.06310.9567 ±0.0000 27 12.7

AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. 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, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-06.

Middlebury 2014 trainingH left

AbsRel / delta1training data not stated: possibly seen

RundtypeAbsRel ↓δ1 ↑Δ AbsRel vs fp32 originalimagesms / image
Original fp32fp320.02540.9954 reference 15 129
Original fp16bf160.02530.9954 +0.0001 15 82.6
Portable fp32not measured
Portablefp160.02540.9954 ±0.0000 15 27.8
Kernelsfp160.02540.9954 ±0.0000 15 15.5

AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. 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, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 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| <= 0.005 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.5 % and p99 <= 2 % of the reference's median |value| (absolute error of the output depth).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.005 (relative)not recorded
GPU default dtypes, mean |diff|≤ 0.005 (relative)0.00032 (relative)
GPU default dtypes, p99 |diff|≤ 0.02 (relative)0.00078 (relative)

Recorded for da3mono-large on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12, with these kernels: dpt_fusedfastbb

Details

Depth Anything 3 (Lin et al. 2025, ByteDance): monocular depth from a DINOv2 ViT with DPT heads; relative (Mono), metric (Metric), any-view depth with camera (Large, Giant) and Giant aligned to Metric (Nested).

Input / output
Input: one RGB image. Output: a depth or inverse-depth map.
What the numbers mean, and which models need the camera focal length

kind (a card field, result.kind) says what the numbers mean: metric = metres; scale_invariant, affine_invariant, affine_invariant_log (log depth) and affine_invariant_inverse are relative (unknown scale or shift, not metres). Metric models that use the focal length take model(image, intrinsics=K) (K [3, 3], pixels of the input): metric3d-vit-large and da3metric-large need it for correct metres (da3metric-large: metres = output x focal / 300; without intrinsics= a 300 px focal length is assumed, wrong for a real camera). unidepth-v2-vitl14, unik3d-vitl, depth-pro and MoGe estimate the camera themselves (given intrinsics= replace fx); zoedepth-nk and the relative models (da3mono-large too) take none. Normals-only models (kind none) return result.normals [B, H, W, 3] (unit vectors, camera frame: x right, y down, z forward).

Input size
any H x W >= 8, resized or padded inside the model (ViT patch 14); tested up to 3000x4000 (1,896 MiB allocated)
Options
  • long_side=504>= 14· changes the output
Outputs
depth, confidence, intrinsics, mask, sky(on request)
Checkpoint
arch="mono"
CUDA kernels
  • fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbones
  • dpt_fusedfused fp16 matrix multiplies and convolutions for DINOv2 backbones and DPT depth heads
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

ByteDance's Depth Anything 3: depth from one image, with relative, metric and multi-view versions. The licence differs by variant, see the table.

Commercial use: da3mono-large, da3metric-large: yes; da3-large-1.1, da3-giant-1.1, da3nested-giant-large-1.1: no (weights: CC-BY-NC-4.0).

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
da3mono-largesynthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
127
85.2
24.1
14.9
21.2
14.4
2419 / 2450
2997 / 3052
814 / 868
805 / 868
792 / 828
1463 / 1520
da3metric-largesynthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
128
85.1
24.3
15.0
21.2
14.4
2418 / 2450
2996 / 3050
811 / 868
802 / 848
792 / 828
1465 / 1502
da3-large-1.1synthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
208
166
30.0
17.3
26.5
13.8
2797 / 2872
3374 / 3472
988 / 1088
891 / 950
841 / 868
1562 / 1622
da3-giant-1.1synthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
410
227
71.5
45.2
64.3
37.2
6428 / 6496
8590 / 8718
2634 / 2728
2535 / 2590
2487 / 2528
4795 / 4842
da3nested-giant-large-1.1synthetic (994x1482)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
533
308
94.9
59.7
85.4
51.9
7702 / 7760
10332 / 10402
3271 / 3374
3179 / 3238
3124 / 3194
6152 / 6226
Details
  • Original code: https://github.com/ByteDance-Seed/Depth-Anything-3 at 3d835ec
  • Code licence: Apache-2.0 (permissive); files: LICENSE
  • Weights licence: Apache-2.0
    • da3-large-1.1: CC-BY-NC-4.0
    • da3-giant-1.1: CC-BY-NC-4.0
    • da3nested-giant-large-1.1: CC-BY-NC-4.0
  • Licence evidence: https://github.com/ByteDance-Seed/Depth-Anything-3/blob/main/README.md (Model Cards table) and the Hugging Face model cards (DA3-LARGE-1.1's card says Apache 2.0, the repository table CC BY-NC 4.0; the table is recorded)
  • Training data: Public academic datasets only (model cards); teacher-student training on synthetic and real multi-view data per the paper
  • Licence note: The Mono and Metric weights are Apache-2.0 but their training data terms are unchecked (a note only); the other variants' weights are CC-BY-NC-4.0 (no).
  • Options: long_side = 504 (>= 14; also process_res)
  • Kernel libraries: fastbb, dpt_fused
  • Variants: da3mono-large, da3metric-large, da3-large-1.1, da3-giant-1.1, da3nested-giant-large-1.1
  • Outputs: depth, confidence, intrinsics, mask, sky (in extras); kind: metric
  • Input size: any H x W >= 8, resized or padded inside the model (ViT patch 14)
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): da3mono-large 3000x4000 (1.9 GiB peak allocated, 2.1 reserved); da3metric-large 3000x4000 (1.9 GiB peak allocated, 2.1 reserved); da3-large-1.1 3000x4000 (1.9 GiB peak allocated, 2.2 reserved); da3-giant-1.1 3000x4000 (5.1 GiB peak allocated, 5.4 reserved); da3nested-giant-large-1.1 3000x4000 (6.3 GiB peak allocated, 6.7 reserved)
  • da3mono-large 994x1482, first call: portable + torch.compile 25 s (cold Inductor cache), TensorRT fp16 52 s (ONNX + engine).
  • da3metric-large 994x1482, first call: portable + torch.compile 27 s (cold Inductor cache), TensorRT fp16 58 s (ONNX + engine).
  • da3-large-1.1 994x1482, first call: portable + torch.compile 45 s (cold Inductor cache), TensorRT fp16 63 s (ONNX + engine).
  • da3-giant-1.1 994x1482, first call: portable + torch.compile 47 s (cold Inductor cache), TensorRT fp16 84 s (ONNX + engine).
  • da3nested-giant-large-1.1 994x1482, first call: portable + torch.compile 60 s (cold Inductor cache), TensorRT fp16 154 s (ONNX + engine).
  • da3mono-large original (fp32) 994x1482: float32 (the authors' autocast off)
  • da3mono-large original 994x1482: original default (api.py): torch.autocast(bfloat16)
  • da3metric-large original (fp32) 994x1482: float32 (the authors' autocast off)
  • da3metric-large original 994x1482: original default (api.py): torch.autocast(bfloat16)
  • da3-large-1.1 original (fp32) 994x1482: float32 (the authors' autocast off)
  • da3-large-1.1 original 994x1482: original default (api.py): torch.autocast(bfloat16)
  • da3mono-large load / first call: original 2.96 s / 0.282 s, portable 2.4 s / 0.209 s (504x504), kernels 2.23 s / 0.17 s (504x504)
  • da3metric-large load / first call: original 2.7 s / 0.283 s, portable 4.84 s / 0.243 s (518x770), kernels 2.34 s / 0.215 s (518x770)
  • da3-large-1.1 load / first call: original 3.46 s / 0.45 s, portable 1.59 s / 0.408 s (518x770), kernels 2.16 s / 0.26 s (518x770)
  • da3-giant-1.1 load / first call: original 12.39 s / 0.684 s, portable 7.21 s / 0.357 s (518x770), kernels 5.38 s / 0.369 s (518x770)
  • da3nested-giant-large-1.1 load / first call: original 28.67 s / 1.626 s, portable 18.24 s / 0.455 s (518x770), kernels 7.35 s / 0.184 s (518x770)
  • 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

AbsRel / delta1 after a least-squares scale + shift fit (in disparity space for inverse-depth models), mean over images. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portableportable fp32kernelsnote
da3mono-largeNYUv2 test, AbsRel / delta10.0402 / 0.98210.0402 / 0.98200.0402 / 0.98210.0402 / 0.98210.0402 / 0.9821
da3mono-largeiBims-1, AbsRel / delta10.0311 / 0.98720.0311 / 0.98720.0311 / 0.98720.0311 / 0.98720.0311 / 0.9873
da3mono-largeETH3D two-view train left, AbsRel / delta10.0631 / 0.95670.0629 / 0.95830.0630 / 0.95680.0631 / 0.9567possibly in the training data
da3mono-largeMiddlebury 2014 trainingH left, AbsRel / delta10.0254 / 0.99540.0253 / 0.99540.0254 / 0.99540.0254 / 0.9954possibly in the training data
da3metric-largeNYUv2 test, AbsRel / delta10.0398 / 0.98200.0398 / 0.98200.0398 / 0.98200.0398 / 0.98200.0398 / 0.9820
da3metric-largeiBims-1, AbsRel / delta10.0328 / 0.98640.0328 / 0.98630.0328 / 0.98640.0328 / 0.98640.0328 / 0.9864
da3metric-largeETH3D two-view train left, AbsRel / delta10.0716 / 0.93430.0716 / 0.93430.0716 / 0.93420.0717 / 0.9342possibly in the training data
da3metric-largeMiddlebury 2014 trainingH left, AbsRel / delta10.0274 / 0.99490.0275 / 0.99490.0274 / 0.99490.0274 / 0.9949possibly in the training data
da3-large-1.1NYUv2 test, AbsRel / delta10.0468 / 0.97520.0468 / 0.97520.0468 / 0.97520.0468 / 0.97520.0468 / 0.9752
da3-large-1.1iBims-1, AbsRel / delta10.0413 / 0.97630.0413 / 0.97640.0413 / 0.97630.0413 / 0.97630.0413 / 0.9764
da3-large-1.1ETH3D two-view train left, AbsRel / delta10.0867 / 0.91770.0869 / 0.91760.0868 / 0.91770.0867 / 0.91770.0867 / 0.9177possibly in the training data
da3-large-1.1Middlebury 2014 trainingH left, AbsRel / delta10.0418 / 0.99010.0419 / 0.98990.0418 / 0.99010.0418 / 0.99010.0419 / 0.9900possibly in the training data
da3-giant-1.1NYUv2 test, AbsRel / delta10.0397 / 0.97940.0397 / 0.97940.0397 / 0.97940.0397 / 0.97940.0397 / 0.9794
da3-giant-1.1iBims-1, AbsRel / delta10.0318 / 0.98190.0317 / 0.98200.0318 / 0.98200.0318 / 0.98190.0318 / 0.9819
da3-giant-1.1ETH3D two-view train left, AbsRel / delta10.0705 / 0.94160.0704 / 0.94150.0704 / 0.94150.0705 / 0.94160.0704 / 0.9420possibly in the training data
da3-giant-1.1Middlebury 2014 trainingH left, AbsRel / delta10.0284 / 0.99300.0284 / 0.99300.0284 / 0.99300.0284 / 0.99300.0284 / 0.9930possibly in the training data
da3nested-giant-large-1.1NYUv2 test, AbsRel / delta10.0387 / 0.98090.0387 / 0.98100.0387 / 0.98090.0387 / 0.98090.0387 / 0.9809
da3nested-giant-large-1.1iBims-1, AbsRel / delta10.0320 / 0.98250.0320 / 0.98260.0320 / 0.98240.0320 / 0.98250.0320 / 0.9825
da3nested-giant-large-1.1ETH3D two-view train left, AbsRel / delta10.0761 / 0.93610.0762 / 0.93460.0761 / 0.93620.0760 / 0.9362possibly in the training data
da3nested-giant-large-1.1Middlebury 2014 trainingH left, AbsRel / delta10.0298 / 0.99250.0299 / 0.99250.0298 / 0.99250.0298 / 0.9925possibly in the training data

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05 to 2026-11-06; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10 / 6d3bab64, 2026-11-05 to 2026-11-17; kernels: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05 to 2026-11-06.

Parity

CPU, float32: max |diff| <= 0.005 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.5 % and p99 <= 2 % of the reference's median |value| (absolute error of the output depth).

variantkernels error vs original, mean / p99GPU, torchkernels used
da3mono-large0.00032 / 0.00078RTX 5060 Ti, torch 2.14.1+cu130dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120
da3metric-large0.00196 / 0.00259RTX 5060 Ti, torch 2.14.1+cu130dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120
da3-large-1.10.00011 / 0.00031RTX 5060 Ti, torch 2.14.1+cu130dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120
da3-giant-1.10.00014 / 0.00037RTX 5060 Ti, torch 2.14.1+cu130dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d:sm_120
da3nested-giant-large-1.10.00151 / 0.00184RTX 5060 Ti, torch 2.14.1+cu130dpt_fused@48e0065bb7cf:sm_120, fastbb@7d058322634d: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).

Input
Output
A cup of coffee, input
Input · 100%

Input: A cup of coffee, Rachel Michetti (CC0).