Models · Optical flow

GMFlow

GMFlow (2022; the flow models of UniMatch, 2023): optical flow as one global matching step between Transformer-enhanced features of both images; variants add a second matching scale and refinement steps. MIT code; the authors' weights carry no separate licence.

experimental Fork of autonomousvision/unimatch@943a990 · backends: portable, kernels

Checkpoint

Paper 2021 Added in 0.2.0 Evaluated with 0.2.0

Portable, RTX 5060 Ti 16 GB 556 ms 1080x1920, fp16
vs the faster original 1.3× original fp16: 750 ms
Peak memory 4,496 MiB original: 8,360 MiB
Accuracy vs original fp32 +0.001 portable EPE, KITTI 2015 train (10.347 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("gmflow-scale1-things")  # kernels on a supported GPU, else portable PyTorch
result = model(frame1, frame2)
result.flow
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench gmflow-scale1-things

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 UniMatch ({'num_scales': 1, 'upsample_factor': 8, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2], 'corr_radius_list': [-1], 'prop_radius_list': [-1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 16fp32 12450.60×8,384 MiB/8,458 MiB1.51.69
Original fp16torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 1, 'upsample_factor': 8, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2], 'corr_radius_list': [-1], 'prop_radius_list': [-1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 16fp16 750–8,360 MiB/8,468 MiB0.10.94
Portable fp32fp32not measured ––––
Portablefp16 5561.3×4,496 MiB/4,642 MiB0.20.56
Kernelsnot measured ––––
Portable + torch.compile 3822.0×4,293 MiB/4,558 MiB35 (build)–
TensorRT fp16fp16 1874.0×4,172 MiBenginenot comparable69 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-11 to 2026-11-15. 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, 2026-11-14Originalfp325143 12.8 GiB process–1080x1920
Portablefp1610676 12.3 GiB processmatches1080x1920
Portablefp323269 11.6 GiB processmatches1080x1920
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
MIT AND Apache-2.0 AND BSD-3-Clause LICENSELICENSE-Apache-2.0LICENSE-RAFT
Weights
MIT AND 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

gmflow-scale1-things-e9887eda.pth 18.8 MB· hosted on s3.eu-central-1.amazonaws.com

fastforks downloads and checks these files on first use (fastforks download gmflow-scale1-things).

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 %

RundtypeEPE ↓Fl-all % ↓Δ EPE vs fp32 originalpairsms / pair
Original fp32fp3210.34734.31 reference 200 90.9
Original fp16fp1610.34734.30 −0.001 200 53.5
Portable fp32fp3210.34734.30 ±0.000 200 77.8
Portablefp1610.34634.30 +0.001 200 56.1
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| <= 0.0001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0031 and p99 <= 0.021 (absolute error of the output flow).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.0001 not recorded
GPU default dtypes, mean |diff|≤ 0.0031 0.00198
GPU default dtypes, p99 |diff|≤ 0.021 0.00808

Recorded for gmflow-scale1-things on RTX 5060 Ti, torch 2.10.0+cu130, 2026-11-17, with these kernels: flow_upsample

Details

GMFlow (Xu, Zhang, Cai, Rezatofighi & Tao 2022) in its UniMatch release (Xu et al. 2023): optical flow as global matching. A CNN gives features at 1/8 resolution, a Transformer with self- and cross-attention (shifted windows) makes the two images' features aware of each other, the flow is the expected position of the best match over the whole image, and a self-attention layer spreads it into occluded areas. `scale2` variants repeat this at 1/4 resolution with local matching; `regrefine6` variants add 6 RAFT-like local refinements.

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 >= 64, padded to a multiple of 16; tested up to 1024x2048 (5,554 MiB allocated)
Outputs
flow
CUDA kernels
  • flow_upsample
Apple (MPS)
fp32, fp16

Limitations

  • TensorRT builds at 1080x1920 fail for scale2 variants with a Myelin CHECK(is_const()) error on RTX 5060 Ti; scale1 variants build.
Full reference, generated from the repository

GMFlow (2022; the flow models of UniMatch, 2023): optical flow as one global matching step between Transformer-enhanced features of both images; variants add a second matching scale and refinement steps. MIT code; the authors' weights carry no separate licence.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
gmflow-scale1-thingssynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
1245
750
556
–
382
187
8384 / 8458
8360 / 8468
4496 / 4642
–
4293 / 4558
4333 / 4390
gmflow-scale1-mixdatasynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
1246
750
556
–
382
187
8384 / 8538
8360 / 8448
4496 / 4642
–
4293 / 4558
4333 / 4390
gmflow-scale2-thingssynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2537
1599
1139
–
765
planned for 0.2.1
8512 / 8742
8425 / 8552
6017 / 6224
–
10865 / 10904
–
gmflow-scale2-sintelsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2541
1607
1139
–
765
planned for 0.2.1
8512 / 8742
8425 / 8552
6017 / 6224
–
10865 / 10904
–
gmflow-scale2-mixdatasynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
2544
1608
1139
–
765
planned for 0.2.1
8512 / 8722
8425 / 8572
6017 / 6224
–
10865 / 10904
–
gmflow-scale2-regrefine6-thingssynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
3230
2175
1665
–
1116
planned for 0.2.1
8522 / 8730
8537 / 8822
6080 / 6230
–
10875 / 10972
–
gmflow-scale2-regrefine6-sintelftsynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
3255
2184
1674
–
1116
planned for 0.2.1
8522 / 8730
8537 / 8822
6080 / 6250
–
10875 / 10972
–
gmflow-scale2-regrefine6-kittisynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
3255
2188
1678
–
1116
planned for 0.2.1
8522 / 8710
8537 / 8822
6080 / 6230
–
10875 / 10972
–
gmflow-scale2-regrefine6-mixdatasynthetic (1080x1920)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
3254
2188
1675
–
1116
planned for 0.2.1
8522 / 8710
8537 / 8822
6080 / 6250
–
10875 / 10972
–
Details
  • Original code: https://github.com/autonomousvision/unimatch at 943a99006950
  • Code licence: MIT AND Apache-2.0 AND BSD-3-Clause (permissive); files: LICENSE, LICENSE-Apache-2.0, LICENSE-RAFT
  • Weights licence: MIT AND Apache-2.0 AND BSD-3-Clause (assumed from code)
  • Licence evidence: https://github.com/autonomousvision/unimatch/blob/943a99006950ab6c84ea72a72beb147fdd03c19a/MODEL_ZOO.md
  • Training data: FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa) (FlyingChairs, then FlyingThings3D (part of SceneFlow))
    • gmflow-scale1-mixdata: MODEL_ZOO.md: "trained on several mixed public datasets" (not listed)
    • gmflow-scale2-sintel: FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), MPI Sintel optical flow, KITTI stereo / optical flow 2012 / 2015, HCI HD1K optical flow benchmark (FlyingChairs, FlyingThings3D, then upstream's 'sintel' stage: Sintel, KITTI 2015, HD1K and FlyingThings3D)
    • gmflow-scale2-mixdata: MODEL_ZOO.md: "trained on several mixed public datasets" (not listed)
    • gmflow-scale2-regrefine6-sintelft: FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), MPI Sintel optical flow, KITTI stereo / optical flow 2012 / 2015, HCI HD1K optical flow benchmark (FlyingChairs, FlyingThings3D, the 'sintel' stage (Sintel, KITTI 2015, HD1K, FlyingThings3D), then fine-tuned on Sintel)
    • gmflow-scale2-regrefine6-kitti: FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa), KITTI stereo / optical flow 2012 / 2015, not stated by the authors (FlyingChairs, FlyingThings3D, Virtual KITTI 2 (no dataset entry here yet, hence 'unstated'), then KITTI 2012 + 2015 (upstream scripts/gmflow_scale2_regrefine6_train.sh, stages vkitti2 and kitti_mix))
    • gmflow-scale2-regrefine6-mixdata: MODEL_ZOO.md: "trained on several mixed public datasets" (not listed)
  • Kernel libraries: flow_upsample
  • Variants: gmflow-scale1-things, gmflow-scale1-mixdata, gmflow-scale2-things, gmflow-scale2-sintel, gmflow-scale2-mixdata, gmflow-scale2-regrefine6-things, gmflow-scale2-regrefine6-sintelft, gmflow-scale2-regrefine6-kitti, gmflow-scale2-regrefine6-mixdata
  • Not in this release (planned for 0.2.1): gmflow-scale2-things TensorRT export: the TensorRT 10 engine build fails (Myelin error on the attention graph); gmflow-scale2-sintel TensorRT export: the TensorRT 10 engine build fails (Myelin error on the attention graph); gmflow-scale2-mixdata TensorRT export: the TensorRT 10 engine build fails (Myelin error on the attention graph); gmflow-scale2-regrefine6-things TensorRT export: the TensorRT 10 engine build fails (Myelin error on the attention graph); gmflow-scale2-regrefine6-sintelft TensorRT export: the TensorRT 10 engine build fails (Myelin error on the attention graph); gmflow-scale2-regrefine6-kitti TensorRT export: the TensorRT 10 engine build fails (Myelin error on the attention graph); gmflow-scale2-regrefine6-mixdata TensorRT export: the TensorRT 10 engine build fails (Myelin error on the attention graph)
  • Outputs: flow
  • Input size:
    • gmflow-scale1-things, gmflow-scale1-mixdata: any H x W >= 64, padded to a multiple of 16
    • gmflow-scale2-things, gmflow-scale2-sintel, gmflow-scale2-mixdata, gmflow-scale2-regrefine6-things, gmflow-scale2-regrefine6-sintelft, gmflow-scale2-regrefine6-kitti, gmflow-scale2-regrefine6-mixdata: any H x W >= 64, padded to a multiple of 32
  • Tested up to (portable, fp16, RTX 5060 Ti 16 GB): gmflow-scale1-things 1024x2048 (5.4 GiB peak allocated, 6.0 reserved); gmflow-scale1-mixdata 1024x2048 (5.4 GiB peak allocated, 6.0 reserved); gmflow-scale2-things 1024x2048 (6.3 GiB peak allocated, 13.1 reserved); gmflow-scale2-sintel 1024x2048 (6.3 GiB peak allocated, 13.1 reserved); gmflow-scale2-mixdata 1024x2048 (6.3 GiB peak allocated, 13.1 reserved); gmflow-scale2-regrefine6-things 1024x2048 (6.3 GiB peak allocated, 13.1 reserved); gmflow-scale2-regrefine6-sintelft 1024x2048 (6.3 GiB peak allocated, 13.1 reserved); gmflow-scale2-regrefine6-kitti 1024x2048 (6.3 GiB peak allocated, 13.1 reserved); gmflow-scale2-regrefine6-mixdata 1024x2048 (6.3 GiB peak allocated, 13.1 reserved)
  • gmflow-scale1-things 1080x1920, first call: portable + torch.compile 35 s (cold Inductor cache), TensorRT fp16 69 s (ONNX + engine).
  • gmflow-scale1-mixdata 1080x1920, first call: portable + torch.compile 29 s (cold Inductor cache), TensorRT fp16 67 s (ONNX + engine).
  • gmflow-scale2-things 1080x1920, first call: portable + torch.compile 47 s (cold Inductor cache).
  • gmflow-scale2-sintel 1080x1920, first call: portable + torch.compile 47 s (cold Inductor cache).
  • gmflow-scale2-mixdata 1080x1920, first call: portable + torch.compile 47 s (cold Inductor cache).
  • gmflow-scale2-regrefine6-things 1080x1920, first call: portable + torch.compile 58 s (cold Inductor cache).
  • gmflow-scale2-regrefine6-sintelft 1080x1920, first call: portable + torch.compile 58 s (cold Inductor cache).
  • gmflow-scale2-regrefine6-kitti 1080x1920, first call: portable + torch.compile 58 s (cold Inductor cache).
  • gmflow-scale2-regrefine6-mixdata 1080x1920, first call: portable + torch.compile 59 s (cold Inductor cache).
  • gmflow-scale1-things original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 1, 'upsample_factor': 8, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2], 'corr_radius_list': [-1], 'prop_radius_list': [-1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 16
  • gmflow-scale1-things original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 1, 'upsample_factor': 8, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2], 'corr_radius_list': [-1], 'prop_radius_list': [-1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 16
  • gmflow-scale1-mixdata original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 1, 'upsample_factor': 8, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2], 'corr_radius_list': [-1], 'prop_radius_list': [-1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 16
  • gmflow-scale1-mixdata original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 1, 'upsample_factor': 8, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2], 'corr_radius_list': [-1], 'prop_radius_list': [-1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 16
  • gmflow-scale2-things original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-things original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-sintel original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-sintel original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-mixdata original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-mixdata original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': False}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 1, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-regrefine6-things original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': True}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 6, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-regrefine6-things original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': True}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 6, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-regrefine6-sintelft original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': True}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 6, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-regrefine6-sintelft original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': True}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 6, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-regrefine6-kitti original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': True}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 6, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-regrefine6-kitti original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': True}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 6, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-regrefine6-mixdata original (fp32) 1080x1920: upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': True}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 6, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale2-regrefine6-mixdata original 1080x1920: torch.autocast(float16) around the call; upstream UniMatch ({'num_scales': 2, 'upsample_factor': 4, 'reg_refine': True}, {'attn_type': 'swin', 'attn_splits_list': [2, 8], 'corr_radius_list': [-1, 4], 'prop_radius_list': [-1, 1], 'num_reg_refine': 6, 'task': 'flow'}), padded to multiples of 32
  • gmflow-scale1-things load / first call: original 0.14 s / 0.942 s, portable 0.16 s / 0.558 s
  • gmflow-scale1-mixdata load / first call: original 0.13 s / 0.76 s, portable 0.16 s / 0.558 s
  • gmflow-scale2-things load / first call: original 0.14 s / 1.61 s, portable 0.17 s / 1.141 s
  • gmflow-scale2-sintel load / first call: original 0.13 s / 1.618 s, portable 0.16 s / 1.141 s
  • gmflow-scale2-mixdata load / first call: original 0.13 s / 1.618 s, portable 0.16 s / 1.141 s
  • gmflow-scale2-regrefine6-things load / first call: original 0.15 s / 2.186 s, portable 0.16 s / 1.666 s
  • gmflow-scale2-regrefine6-sintelft load / first call: original 0.15 s / 2.369 s, portable 0.26 s / 1.676 s
  • gmflow-scale2-regrefine6-kitti load / first call: original 0.15 s / 2.199 s, portable 0.46 s / 1.679 s
  • gmflow-scale2-regrefine6-mixdata load / first call: original 0.14 s / 2.198 s, portable 0.25 s / 1.676 s
  • 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
gmflow-scale1-thingsKITTI 2015 train, EPE / Fl-all %10.347 / 34.3110.347 / 34.3010.346 / 34.3010.347 / 34.30
gmflow-scale1-mixdataKITTI 2015 train, EPE / Fl-all %3.698 / 17.603.698 / 17.613.697 / 17.613.698 / 17.60possibly in the training data
gmflow-scale2-thingsKITTI 2015 train, EPE / Fl-all %6.747 / 22.096.747 / 22.096.747 / 22.096.747 / 22.09
gmflow-scale2-sintelKITTI 2015 train, EPE / Fl-all %1.975 / 7.881.976 / 7.881.975 / 7.881.975 / 7.88trained on this dataset's public images; not a zero-shot result
gmflow-scale2-mixdataKITTI 2015 train, EPE / Fl-all %2.390 / 10.102.390 / 10.102.389 / 10.102.390 / 10.10possibly in the training data
gmflow-scale2-regrefine6-thingsKITTI 2015 train, EPE / Fl-all %4.918 / 16.334.918 / 16.324.917 / 16.324.918 / 16.33
gmflow-scale2-regrefine6-sintelftKITTI 2015 train, EPE / Fl-all %1.153 / 3.921.154 / 3.921.154 / 3.921.153 / 3.92trained on this dataset's public images; not a zero-shot result
gmflow-scale2-regrefine6-kittiKITTI 2015 train, EPE / Fl-all %0.570 / 1.150.570 / 1.150.570 / 1.150.570 / 1.15trained on this dataset's public images; not a zero-shot result
gmflow-scale2-regrefine6-mixdataKITTI 2015 train, EPE / Fl-all %1.315 / 4.641.315 / 4.631.315 / 4.641.314 / 4.64possibly in the training data

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| <= 0.0001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0031 and p99 <= 0.021 (absolute error of the output flow).

variantkernels error vs original, mean / p99GPU, torchkernels used
gmflow-scale1-things0.00198 / 0.00808RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120
gmflow-scale1-mixdata0.002 / 0.00938RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120
gmflow-scale2-things0.00128 / 0.00537RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120
gmflow-scale2-sintel0.00134 / 0.00653RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120
gmflow-scale2-mixdata0.00145 / 0.00581RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120
gmflow-scale2-regrefine6-things0.00098 / 0.00504RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120
gmflow-scale2-regrefine6-sintelft0.00112 / 0.00836RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120
gmflow-scale2-regrefine6-kitti0.00175 / 0.01264RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120
gmflow-scale2-regrefine6-mixdata0.00098 / 0.00564RTX 5060 Ti, torch 2.10.0+cu130flow_upsample@5eca8c1a6d3b:sm_120

Limitations

  • TensorRT builds at 1080x1920 fail for scale2 variants with a Myelin CHECK(is_const()) error on RTX 5060 Ti; scale1 variants build.

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).