flowformer / README.md
Flor's picture
card: update from fastforks 1cbf1671
d45061d verified
|
Raw History Blame Contribute Delete
8.35 kB
metadata
library_name: fastforks
pipeline_tag: image-to-image
license: other
license_name: see-model-card
license_link: >-
  https://github.com/drinkingcoder/FlowFormer-Official/blob/6ba7ea82b45394d3a7a25808399695427c9febd8/README.md#license
tags:
  - flow
  - optical-flow
  - motion-estimation
  - cuda
  - pytorch
  - fastforks
  - fastforks-card

FlowFormer Small (fastforks)

FlowFormer Small (ECCV 2022): convolutional features, a latent cost-memory transformer and recurrent flow decoding. The authors' Small checkpoints, with plain float32 inference.

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/drinkingcoder/FlowFormer-Official at 6ba7ea82b453.

Task: flow. Tier: experimental. Outputs: flow. Backends: portable.

Input: two RGB images (consecutive frames). Output: per-pixel motion from the first to the second (dx, dy in pixels).

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant weights (author link) SHA-256 weights licence commercial use
flowformer-small-chairs@1 flowformer-small-chairs.pth 353914dcbf90 Apache-2.0 AND BSD-3-Clause AND WTFPL (assumed from code) yes
flowformer-small-sintel@1 flowformer-small-sintel.pth 0ed780e00504 Apache-2.0 AND BSD-3-Clause AND WTFPL (assumed from code) yes
flowformer-small-things@1 flowformer-small-things.pth 6bd2b44f3170 Apache-2.0 AND BSD-3-Clause AND WTFPL (assumed from code) yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0 AND BSD-3-Clause AND WTFPL. Licence files: LICENSE, LICENSE-RAFT, LICENSE-GMA.

Code licence: Apache-2.0 AND BSD-3-Clause AND WTFPL; the authors do not state a separate weights licence.

Training data: FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa) (The authors' Small checkpoints follow the chairs, things and sintel training stages; per-variant training data follows configs/small_*.py.).

Licence evidence: https://github.com/drinkingcoder/FlowFormer-Official/blob/6ba7ea82b45394d3a7a25808399695427c9febd8/README.md#license

Input size

any H x W >= 64, padded to a multiple of 8 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): flowformer-small-chairs 1024x1024 (11.2 GiB peak allocated, 12.3 reserved); flowformer-small-sintel 1024x1024 (11.2 GiB peak allocated, 12.3 reserved); flowformer-small-things 1024x1024 (11.2 GiB peak allocated, 12.3 reserved)

Speed and memory

Flow, synthetic

model input time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
flowformer-small-chairs 375x1242 352 303 351 – – – 2382 / 2970 3165 / 3638 –
flowformer-small-sintel 375x1242 352 303 351 – – – 2382 / 2990 3165 / 3638 –
flowformer-small-things 375x1242 352 303 351 – – – 2382 / 3010 3165 / 3638 –

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
flowformer-small-chairs 0.16 (375x1242) 1.31 (375x1242) – 0.37 (375x1242) 0.57 (375x1242) –
flowformer-small-sintel 0.15 (375x1242) 1.30 (375x1242) – 0.30 (375x1242) 0.57 (375x1242) –
flowformer-small-things 0.15 (375x1242) 1.31 (375x1242) – 0.30 (375x1242) 0.57 (375x1242) –

torch.compile and TensorRT

model input time: portable compile TensorRT fp16 kernels preparation: compile (cold first call) ONNX + engine build peak MiB (alloc / reserved): portable compile kernels TensorRT (engine + torch)
flowformer-small-chairs 375x1242 351 ms 250 ms 202 ms – 98.7 s 118 s 3165 / 3638 2438 / 3196 – 2087 / 2138
flowformer-small-sintel 375x1242 351 ms 250 ms 202 ms – 98.9 s 117 s 3165 / 3638 2438 / 3196 – 2087 / 2138
flowformer-small-things 375x1242 351 ms 250 ms 202 ms – 99.3 s 118 s 3165 / 3638 2438 / 3196 – 2087 / 2138

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

On public datasets

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.

variant dataset original fp32 original fp16 portable note
flowformer-small-chairs KITTI 2015 train, EPE / Fl-all % 10.952 / 36.72 11.102 / 36.80 10.952 / 36.72
flowformer-small-things KITTI 2015 train, EPE / Fl-all % 6.884 / 17.92 7.034 / 17.91 6.884 / 17.92
flowformer-small-sintel KITTI 2015 train, EPE / Fl-all % 4.657 / 6.35 4.206 / 6.30 4.657 / 6.35 trained on this dataset's public images; not a zero-shot result

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks a6ea0725+dirty, 2026-11-11.

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.0001 vs the original. GPU, portable float32 and kernels float16 (the defaults): mean |diff| <= 0.0021 and p99 <= 0.013 (absolute error of the output flow).

Limitations

  • Only the Small architecture is included.
  • No custom kernels backend: the time is in cuDNN TF32 convolutions and the float32 aggregation matmul, and lowering precision changes EPE.
  • Portable runs float32 (the validated dtype): 338 ms and 1410 MB at 375x1242 against 303 ms and 2382 MB for the fp16-autocast original (EPE identical to the float32 original); switching the default to fp16 needs a full accuracy re-validation.
  • Full-frame cost memory grows quadratically with input area; upstream tiled evaluation is not included. Full-frame 1080p exceeded the RTX 5060 Ti 16GB memory in portable and original fp32/fp16 runs.

Usage

from fastforks import flow

model = flow.load("flowformer-small-chairs")  # mode="auto": the preferred backend from recorded measurements
result = model("frame1.png", "frame2.png")
result.flow  # [2, H, W] float32 tensor: dx, dy in input pixels
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: flowformer-small-chairs, flowformer-small-sintel, flowformer-small-things.

Kernels backend

No kernels backend: Measured 2026-11-15: time is cuDNN TF32 convolutions plus the float32 aggregation matmul; fp16 aggregation (NaN at scale 1024, worse EPE) and SDPA (slower, EPE changed) were tried and dropped..