--- library_name: "fastforks" pipeline_tag: "image-to-image" license: "other" license_name: "see-model-card" license_link: "https://github.com/pytorch/vision/blob/1f1a920d685d46dfbe187934b785b7ecb7a182f2/torchvision/models/optical_flow/raft.py#L576" tags: - "flow" - "optical-flow" - "motion-estimation" - "cuda" - "pytorch" - "fastforks" - "fastforks-card" --- # RAFT (fastforks) RAFT (Teed & Deng 2020): all-pairs correlation pyramid at 1/8 resolution and a ConvGRU that refines the flow in 12 iterations, upsampled with a learned convex mask (raft-small: bilinear). torchvision's V2 weights (trained from scratch with its optical-flow reference recipe); the network is ported from torchvision (no torchvision dependency). 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: . Kernels for the `kernels` backend: . Original code: at `1f1a920d685d`. Task: flow. Tier: stable. Outputs: `flow`. Backends: `portable`, `kernels`. 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`](https://github.com/kleinicke/fastforks/blob/main/src/fastforks/flow/models/raft/card.toml). | variant | weights (author link) | SHA-256 | weights licence | commercial use | |---|---|---|---|---| | `raft-large@1` | [raft_large_C_T_SKHT_V2-ff5fadd5.pth](https://download.pytorch.org/models/raft_large_C_T_SKHT_V2-ff5fadd5.pth) | `ff5fadd56d26` | BSD-3-Clause (assumed from code) | yes | | `raft-large-kitti@1` | [raft_large_C_T_SKHT_K_V2-b5c70766.pth](https://download.pytorch.org/models/raft_large_C_T_SKHT_K_V2-b5c70766.pth) | `b5c7076664d2` | BSD-3-Clause (assumed from code) | yes | | `raft-large-things@1` | [raft_large_C_T_V2-1bb1363a.pth](https://download.pytorch.org/models/raft_large_C_T_V2-1bb1363a.pth) | `1bb1363a43b4` | BSD-3-Clause (assumed from code) | yes | | `raft-small@1` | [raft_small_C_T_V2-01064c6d.pth](https://download.pytorch.org/models/raft_small_C_T_V2-01064c6d.pth) | `01064c6dba73` | BSD-3-Clause (assumed from code) | yes | ## Licence Original-code strings, recorded as they are (research-use package; check them before any other use). Code: BSD-3-Clause. Licence files: `LICENSE`. Code licence: BSD-3-Clause; the authors do not state a separate weights licence. Training data: FlyingChairs (Freiburg), Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa) (FlyingChairs, then FlyingThings3D (part of SceneFlow)). Licence evidence: ## Input size any H x W >= 128, padded to a multiple of 8 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): `raft-large` 1024x2048 (8.3 GiB peak allocated, 9.0 reserved); `raft-large-kitti` 1024x2048 (8.3 GiB peak allocated, 9.0 reserved); `raft-large-things` 1024x2048 (8.3 GiB peak allocated, 9.0 reserved); `raft-small` 1024x2048 (8.2 GiB peak allocated, 8.7 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 | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | `raft-large` | 1080x1920 | 480 | 567 | 291 | 83.5 | 5.7x | 3.5x | 7164 / 10928 | 5930 / 6058 | 342 / 372 | | `raft-large-kitti` | 1080x1920 | 480 | 567 | 291 | 83.6 | 5.7x | 3.5x | 7164 / 10928 | 5930 / 6058 | 333 / 352 | | `raft-large-things` | 1080x1920 | 481 | 566 | 291 | 83.1 | 5.8x | 3.5x | 7164 / 10928 | 5930 / 6058 | 333 / 352 | | `raft-small` | 1080x1920 | 196 | 395 | 144 | 36.3 | 5.4x | 4.0x | 7032 / 10888 | 5669 / 5706 | 197 / 224 | ### Flow, KITTI 2015, 375x1242 | model | time (ms): portable | kernels | kernels speed-up: vs portable | peak memory (MiB, allocated / reserved): portable | kernels | | --- | --- | --- | --- | --- | --- | | `raft-large` | 45.8 | 21.0 | 2.2x | 511 / 738 | 85 / 112 | | `raft-large-things` | 45.8 | 21.1 | 2.2x | 511 / 738 | 85 / 112 | | `raft-small` | 23.2 | 12.1 | 1.9x | 483 / 508 | 48 / 66 | ### Startup and first call | model | load (s): original | portable | kernels | first call (s): original | portable | kernels | | --- | --- | --- | --- | --- | --- | --- | | `raft-large` | 0.14 | 0.17 | 1.07 | 0.76 | 0.30 | 0.11 | | `raft-large-kitti` | 0.14 | 0.17 | 1.06 | 0.58 | 0.30 | 0.11 | | `raft-large-things` | 0.13 | 0.25 | 1.07 | 0.58 | 0.30 | 0.11 | | `raft-small` | 0.09 | 0.16 | 1.06 | 0.43 | 0.17 | 0.07 | ### 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) | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | `raft-large` | 1080x1920 | 291 ms | 195 ms | fails: TensorRT engine build failed | 83.5 ms | 41.3 s | – | 5930 / 6058 | 5886 / 5924 | 342 / 372 | – | | `raft-large-kitti` | 1080x1920 | 291 ms | 195 ms | fails: TensorRT engine build failed | 83.6 ms | 39.3 s | – | 5930 / 6058 | 5886 / 5924 | 333 / 352 | – | | `raft-large-things` | 1080x1920 | 291 ms | 195 ms | fails: TensorRT engine build failed | 83.1 ms | 39.4 s | – | 5930 / 6058 | 5886 / 5924 | 333 / 352 | – | | `raft-small` | 1080x1920 | 144 ms | 110 ms | fails: unsupported: engine build needs more than the 16 GB of this | 36.3 ms | 37.0 s | – | 5669 / 5706 | 5704 / 5728 | 197 / 224 | – | Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB). - `raft-large-kitti`: same network and settings as `raft-large`, measured once (rows not repeated). Hardware, protocol and caveats: [benchmarks](https://github.com/kleinicke/fastforks/blob/main/docs/benchmarks.md) (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 | kernels | note | |---|---|---|---|---|---|---| | `raft-large` | KITTI 2015 train, EPE / Fl-all % | 1.528 / 5.60 | 1.528 / 5.60 | 1.529 / 5.61 | 1.526 / 5.60 | trained on this dataset's public images; not a zero-shot result | | `raft-large-things` | KITTI 2015 train, EPE / Fl-all % | 4.838 / 16.55 | 4.831 / 16.54 | 4.828 / 16.51 | 4.834 / 16.53 | | | `raft-large-kitti` | KITTI 2015 train, EPE / Fl-all % | 0.614 / 1.44 | 0.614 / 1.44 | 0.614 / 1.44 | 0.614 / 1.44 | trained on this dataset's public images; not a zero-shot result | | `raft-small` | KITTI 2015 train, EPE / Fl-all % | 7.854 / 25.91 | 7.838 / 25.92 | 7.846 / 25.92 | 7.834 / 25.92 | | Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks f2992269+uncommitted, 2026-11-09. ### Parity with the original code Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.00012 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.0019 and p99 <= 0.0091 (absolute error of the output `flow`). ## Usage ```python from fastforks import flow model = flow.load("raft-large") # 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 ). Other variants: `raft-large`, `raft-large-kitti`, `raft-large-things`, `raft-small`. ## Kernels backend `flow.load("raft-large", mode="kernels")` runs our CUDA kernels (`raft_flow`, `cutlass_conv`, `cudnn_hacc`). It needs a prebuilt kernel for your GPU architecture (downloaded from on first use) and a recent NVIDIA driver; if there is none it raises `BackendUnavailableError` with the reason and never falls back silently. `mode="auto"` picks `kernels` when it can run and logs the choice.