Models · Multi-view depth

Pi3

π3 (Pi3, yyfz): cameras, depth and 3D points from many unposed images in one pass, with no reference view. Non-commercial weights.

experimental Fork of yyfz/Pi3@9fa3ddb · backends: portable, kernels

Checkpoint pi3

Paper 2025 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 559 ms 336x518, fp16
vs the faster original 1.9× original bf16: 1057 ms
Peak memory 2,367 MiB original: 5,926 MiB
Accuracy vs original fp32 ±0.0000 kernels AbsRel, DTU scans 1 (0.0070 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("pi3")  # kernels on a supported GPU, else portable PyTorch
result = model(images)
result.depth  # also: confidence, world_points, cam_to_world
result.backend  # what ran: mode, device, dtype, kernels

The weights are non-commercial: fastforks downloads them only with FASTFORKS_NONCOMMERCIAL=1 set.

fastforks bench pi3

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 (own default)the authors' mixed-precision option (bfloat16); yyfz/Pi3@9fa3ddb3f8d5 Pi3.forward, PIXEL_LIMIT 255000bf16 1057–5,926 MiB/6,016 MiB9.11.15
Portable fp32fp32not measured ––––
Portablefp16 8211.3×2,360 MiB/2,416 MiB3.51.28
Kernelsfp16 5591.9×2,367 MiB/2,456 MiB3.80.70
Portable + torch.compile 7521.4×2,205 MiB/2,334 MiB56 (build)–
TensorRTplanned for 0.2.1fails––––

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-13 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, 2026-11-13Originalfp3210379 6,238 MiB process–336x518
Portablefp168341 6,510 MiB processmatches336x518
Portablefp3210298 6,201 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

Non-commercial only

Code
BSD-3-Clause LICENSE
Weights
CC-BY-NC-4.0 source
Training data
Per the paper (arXiv 2507.13347): a mixture of public synthetic and real multi-view datasets; DINOv2 initialisation. The README restricts the weights to non-commercial use because of the training datasets.
  • weights: CC-BY-NC-4.0
  • The Pi3 code is BSD-3-Clause, but the README licenses the Pi3 weights CC-BY-NC-4.0 (non-commercial); the Hugging Face card's bsd-2-clause tag is overridden by that statement.

Our reading of the licences, not legal advice.

Weights

model.safetensors 3.8 GB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download pi3). Non-commercial weights need FASTFORKS_NONCOMMERCIAL=1.

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 fp32fp320.007099.94 reference 5 3605
Original fp16bf160.007199.94 −0.0001 5 1022
Portable fp32fp320.007099.94 ±0.0000 5 3602
Portablefp160.007099.94 ±0.0000 5 823
Kernelsfp160.007099.94 ±0.0000 5 562

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 0.00107
GPU default dtypes, p99 |diff|≤ 0.25 0.00237

Recorded for pi3 on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-13, with these kernels: fastbb

Details

π³, permutation-equivariant visual geometry learning (Wang et al. 2025): cameras, depth and point maps of every view from unposed images in one forward pass, with no reference view. A DINOv2 ViT-L/14 encoder feeds 36 alternating frame / global attention blocks (q/k norm, 2-D RoPE); three small transformer decoders give local point maps, confidences and camera poses.

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 (4,896 MiB allocated)
Options
  • pixel_limit=255000>= 196· changes the output
Outputs
depth, confidence, world_points, cam_to_world
CUDA kernels
  • fastbbfast backbone kernels: flash attention, window attention, instance norm and rotary embeddings for ViT and similar backbones
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

π3 (Pi3, yyfz): cameras, depth and 3D points from many unposed images in one pass, with no reference view. Non-commercial weights.

Commercial use: no (weights: CC-BY-NC-4.0).

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
pi3synthetic (336x518)original
portable
kernels
portable + torch.compile
TensorRT fp16
1057
821
559
752
planned for 0.2.1
5926 / 6016
2360 / 2416
2367 / 2456
2205 / 2334
–
Details
  • Original code: https://github.com/yyfz/Pi3 at 9fa3ddb
  • Code licence: BSD-3-Clause (permissive); files: LICENSE
  • Weights licence: CC-BY-NC-4.0
  • Licence evidence: https://github.com/yyfz/Pi3/blob/9fa3ddb/LICENSE and https://github.com/yyfz/Pi3/blob/9fa3ddb/README.md#-license
  • Training data: Per the paper (arXiv 2507.13347): a mixture of public synthetic and real multi-view datasets; DINOv2 initialisation. The README restricts the weights to non-commercial use because of the training datasets.
  • Licence note: The Pi3 code is BSD-3-Clause, but the README licenses the Pi3 weights CC-BY-NC-4.0 (non-commercial); the Hugging Face card's bsd-2-clause tag is overridden by that statement.
  • Options: pixel_limit = 255000 (>= 196)
  • Kernel libraries: fastbb
  • Variants: pi3
  • Not in this release (planned for 0.2.1): pi3 ONNX / TensorRT export: the ONNX converter fails on the exported program (ConversionError, step 3/3)
  • Outputs: depth, confidence, world_points, 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): pi3 3000x4000 (4.8 GiB peak allocated, 6.5 reserved)
  • pi3 336x518, first call: portable + torch.compile 56 s (cold Inductor cache).
  • pi3 original 336x518: the authors' mixed-precision option (bfloat16); yyfz/Pi3@9fa3ddb3f8d5 Pi3.forward, PIXEL_LIMIT 255000
  • pi3 load / first call: original 10.64 s / 4.059 s (336x518), portable 3.46 s / 1.284 s (336x518), kernels 3.78 s / 0.7 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
pi3DTU scans 1, 9, 24, 33, 75, view 24 + 7 sources, AbsRel / delta1 %0.0070 / 99.940.0071 / 99.940.0070 / 99.940.0070 / 99.940.0070 / 99.94

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
pi30.00107 / 0.00237RTX 5060 Ti, torch 2.14.1+cu130fastbb@7d058322634d:sm_120

Example output

No example rendered for this checkpoint yet.