Models · Optical flow

FlowSeek

FlowSeek (University of Bologna, 2025), size T: SEA-RAFT optical flow that also uses a frozen Depth Anything V2 Small: its features help the matching and its depth tells which flow a camera motion would cause. Apache-2.0 code; the authors' weights carry no separate licence.

experimental Fork of mattpoggi/flowseek@ba64073 · backends: portable, kernels

Checkpoint

Paper 2025 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 327 ms 1080x1920, fp16
vs the faster original 1.4× original fp16: 457 ms
Peak memory 2,027 MiB original: 8,693 MiB
Accuracy vs original fp32 −0.088 portable EPE, KITTI 2015 train (1.260 original)

Use it

pip install fastforks

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

from fastforks import flow

model = flow.load("flowseek-t")  # kernels on a supported GPU, else portable PyTorch
result = model(frame1, frame2)
result.flow
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench flowseek-t

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 fp32upstream FlowSeek (flowseek-T.json), iters=4fp32 6500.70×9,902 MiB/10,052 MiB2.61.03
Original fp16torch.autocast(float16) around the call; upstream FlowSeek (flowseek-T.json), iters=4fp16 457–8,693 MiB/9,880 MiB0.90.63
Portable fp32fp32not measured ––––
Portablefp16 3951.2×8,671 MiB/8,712 MiB0.20.40
Kernelson-demand correlation; code snapshot was measured before a style-only constant rename; 1080p improves speed and memory, KITTI-sized runtime is similar to portablefp16 3271.4×2,027 MiB/2,192 MiB0.30.33
Portable + torch.compile 2581.8×6,126 MiB/6,212 MiB54 (build)–
TensorRT fp16fp16 1682.7×10.8 GiBenginenot comparable107 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-11 to 2026-11-17. 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, 2026-11-14Originalfp32out of memory–1080x1920
Portablefp16out of memory–1080x1920
Portablefp32out of memory–1080x1920
Kernelsfp32not implemented–1080x1920

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 AND BSD-3-Clause LICENSELICENSE-BSD-3-Clause
Weights
Apache-2.0 AND BSD-3-Clause (assumed from code)(not stated by the authors; assumed to follow the code) source
Training data
  • weights licence not stated by the original authors; assumed to be the code licence

Our reading of the licences, not legal advice.

Weights

flowseek_T_TartanCT_TSKH.pth 171 MB· hosted on Google Drive

fastforks downloads and checks these files on first use (fastforks download flowseek-t).

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.

KITTI 2015 train

EPE / Fl-all %trained on this dataset’s public images: not zero-shot

RundtypeEPE ↓Fl-all % ↓Δ EPE vs fp32 originalpairsms / pair
Original fp32fp321.2603.88 reference 200 149
Original fp16fp161.3604.35 −0.099 200 84.8
Portable fp32fp321.2593.88 +0.001 200 154
Portablefp161.3484.33 −0.088 200 81.0
Kernelsnot measured

Frames _10 -> _11 at full size against the flow_occ ground truth; EPE = mean of the per-pair mean end-point error, Fl-all = share of all valid pixels with an error > 3 px and > 5 % of the GT magnitude. Original = the authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-11; portable fp32: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17; portable: RTX 5060 Ti, torch 2.14.1+cu130, fastforks c4a67979, 2026-11-11.

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| <= 0.001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0077 and p99 <= 0.041 (absolute error of the output flow).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.001 not recorded
GPU default dtypes, mean |diff|≤ 0.0077 0.00508
GPU default dtypes, p99 |diff|≤ 0.041 0.0272

Recorded for flowseek-t on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-17, with these kernels: raft_flow

Details

FlowSeek (Poggi & Tosi 2025), size T: SEA-RAFT (ResNet18 encoders, initial flow regression, ConvNeXt update block, 4 iterations) helped by a frozen Depth Anything V2 Small. Its features join the matching features, and its relative depth gives 8 motion bases (the flow fields that camera translations and rotations would cause), which a third encoder turns into extra context. Each image is resized to 518 x 518 for Depth Anything, whatever the input size. The authors' released weights; the network is ported from mattpoggi/flowseek. `iters=12` is the paper's FlowSeek (S). The larger M and L sizes use Depth Anything V2 Base, whose weights are non-commercial, and are not included.

Input / output
Input: two RGB images (consecutive frames). Output: per-pixel motion from the first to the second (dx, dy in pixels).
Input size
any H x W >= 128, resized or padded inside the model; tested up to 1024x2048 (8,850 MiB allocated)
Options
  • iters=4>= 1· changes the output
Outputs
flow
CUDA kernels
  • raft_flowRAFT optical flow kernels
Apple (MPS)
fp32, fp16
Full reference, generated from the repository

FlowSeek (University of Bologna, 2025), size T: SEA-RAFT optical flow that also uses a frozen Depth Anything V2 Small: its features help the matching and its depth tells which flow a camera motion would cause. Apache-2.0 code; the authors' weights carry no separate licence.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
flowseek-tsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
650
457
395
327
258
168
9902 / 10052
8693 / 9880
8671 / 8712
2027 / 2192
6126 / 6212
11328 / 11350
flowseek-t-thingssynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
650
458
395
327
258
168
9902 / 10012
8693 / 9820
8671 / 8712
2027 / 2192
6126 / 6212
11330 / 11352
flowseek-t-things-notartansynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
651
458
396
328
258
168
9902 / 10052
8693 / 9820
8671 / 8712
2027 / 2192
6126 / 6212
11328 / 11350
Details
  • Original code: https://github.com/mattpoggi/flowseek at ba6407370bce
  • Code licence: Apache-2.0 AND BSD-3-Clause (permissive); files: LICENSE, LICENSE-BSD-3-Clause
  • Weights licence: Apache-2.0 AND BSD-3-Clause (assumed from code)
  • Licence evidence: https://github.com/mattpoggi/flowseek/blob/ba6407370bceeb96244cad4d60d37489f6eb471e/scripts/get_weights.sh
  • Training data: TartanAir, FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), MPI Sintel optical flow, KITTI stereo / optical flow 2012 / 2015, HCI HD1K optical flow benchmark (TartanAir pretraining, FlyingChairs, FlyingThings3D (part of SceneFlow), then a mix of FlyingThings3D, Sintel, KITTI 2015 and HD1K (TSKH); the Depth Anything V2 Small inside was trained by its authors on synthetic and pseudo-labelled real images)
    • flowseek-t-things: TartanAir, FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa) (TartanAir, FlyingChairs, then FlyingThings3D (part of SceneFlow))
    • flowseek-t-things-notartan: FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa) (FlyingChairs, then FlyingThings3D (part of SceneFlow))
  • Options: iters = 4 (>= 1)
  • Kernel libraries: raft_flow
  • Variants: flowseek-t, flowseek-t-things, flowseek-t-things-notartan
  • Outputs: flow
  • Input size: any H x W >= 128, resized or padded inside the model
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): flowseek-t 1024x2048 (8.6 GiB peak allocated, 10.2 reserved); flowseek-t-things 1024x2048 (8.6 GiB peak allocated, 10.2 reserved); flowseek-t-things-notartan 1024x2048 (8.6 GiB peak allocated, 10.2 reserved)
  • flowseek-t 1080x1920, first call: portable + torch.compile 54 s (cold Inductor cache), TensorRT fp16 107 s (ONNX + engine).
  • flowseek-t-things 1080x1920, first call: portable + torch.compile 55 s (cold Inductor cache), TensorRT fp16 102 s (ONNX + engine).
  • flowseek-t-things-notartan 1080x1920, first call: portable + torch.compile 54 s (cold Inductor cache), TensorRT fp16 98 s (ONNX + engine).
  • flowseek-t original (fp32) 1080x1920: upstream FlowSeek (flowseek-T.json), iters=4
  • flowseek-t original 1080x1920: torch.autocast(float16) around the call; upstream FlowSeek (flowseek-T.json), iters=4
  • flowseek-t kernels 1080x1920: on-demand correlation; code snapshot was measured before a style-only constant rename; 1080p improves speed and memory, KITTI-sized runtime is similar to portable
  • flowseek-t-things original (fp32) 1080x1920: upstream FlowSeek (flowseek-T.json), iters=4
  • flowseek-t-things original 1080x1920: torch.autocast(float16) around the call; upstream FlowSeek (flowseek-T.json), iters=4
  • flowseek-t-things kernels 1080x1920: on-demand correlation; code snapshot was measured before a style-only constant rename; 1080p improves speed and memory, KITTI-sized runtime is similar to portable
  • flowseek-t-things-notartan original (fp32) 1080x1920: upstream FlowSeek (flowseek-T.json), iters=4
  • flowseek-t-things-notartan original 1080x1920: torch.autocast(float16) around the call; upstream FlowSeek (flowseek-T.json), iters=4
  • flowseek-t-things-notartan kernels 1080x1920: on-demand correlation; code snapshot was measured before a style-only constant rename; 1080p improves speed and memory, KITTI-sized runtime is similar to portable
  • flowseek-t load / first call: original 0.87 s / 0.629 s, portable 0.18 s / 0.458 s (375x1242), kernels 0.28 s / 0.382 s (375x1242)
  • flowseek-t-things load / first call: original 0.78 s / 0.459 s, portable 0.18 s / 0.469 s (375x1242), kernels 0.67 s / 0.404 s (375x1242)
  • flowseek-t-things-notartan load / first call: original 0.81 s / 0.459 s, portable 0.18 s / 0.46 s (375x1242), kernels 0.37 s / 0.383 s (375x1242)
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

Frames _10 -> _11 at full size against the flow_occ ground truth; EPE = mean of the per-pair mean end-point error, Fl-all = share of all valid pixels with an error > 3 px and > 5 % of the GT magnitude. Original = the original code named under Details, same checkpoint and protocol.

variantdatasetoriginal fp32original fp16portableportable fp32note
flowseek-tKITTI 2015 train, EPE / Fl-all %1.260 / 3.881.360 / 4.351.348 / 4.331.259 / 3.88trained on this dataset's public images; not a zero-shot result
flowseek-t-thingsKITTI 2015 train, EPE / Fl-all %4.121 / 12.354.153 / 12.254.143 / 12.274.098 / 12.34
flowseek-t-things-notartanKITTI 2015 train, EPE / Fl-all %3.923 / 12.833.904 / 12.973.933 / 12.973.925 / 12.81

Measured on portable: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks c4a67979, 2026-11-11; portable32: NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 6f0a70e1, 2026-11-17.

Parity

CPU, float32: max |diff| <= 0.001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0077 and p99 <= 0.041 (absolute error of the output flow).

variantkernels error vs original, mean / p99GPU, torchkernels used
flowseek-t0.00508 / 0.0272RTX 5060 Ti, torch 2.14.1+cu130raft_flow@a101a33607b6:sm_120
flowseek-t-things0.00427 / 0.0176RTX 5060 Ti, torch 2.14.1+cu130raft_flow@a101a33607b6:sm_120
flowseek-t-things-notartan0.00227 / 0.01045RTX 5060 Ti, torch 2.14.1+cu130raft_flow@a101a33607b6:sm_120

Example output

One fixed input through this checkpoint, the same for every model of the task (portable · float16 · NVIDIA GeForce RTX 5060 Ti).

Frame 1
Frame 2
Output
Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), frame 1
Frame 1 · 100%

Input: Motorcycle, Middlebury 2014 stereo (left and right view, downscaled 4x), Scharstein et al., Middlebury College (research dataset, no licence stated).