LiteAnyStereo V2 (fastforks)
LiteAnyStereo V2: efficient zero-shot stereo with FasterNet-T0 features, a correlation volume and FasterNet 2-D cost aggregation (sizes S / M / L), or a group-wise volume with a selective-GRU update (H).
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/TomTomTommi/LiteAnyStereo at 8c97bd4c4da3.
Author weight repos: https://huggingface.co/tomtomtommi/LiteAnyStereoV2.
Task: stereo. Tier: experimental. Outputs: disparity. Backends: portable, kernels.
Input: a rectified left / right image pair. Output: a disparity map (pixels, left view).
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 |
|---|---|---|---|---|
las2-h@1 |
LAS2_H.pth | 758585a25c3a |
MIT | yes |
las2-l@1 |
LAS2_L.pth | 12f86f758b0a |
MIT | yes |
las2-m@1 |
LAS2_M.pth | c41b61efce9b |
MIT | yes |
las2-s@1 |
LAS2_S.pth | dda5ad549571 |
MIT | yes |
Licence
Original-code strings, recorded as they are (research-use package; check them before any other use). Code: MIT. Files: LICENSE.
Training data: Zero-shot model trained on a mixture of synthetic stereo datasets that includes FSD and Dynamic Replica (non-commercial data terms); the repository does not list the data or its terms.
- Licence note: Trained on a mixture that includes NVIDIA's FSD dataset and Dynamic Replica, whose terms for commercial use of trained models are unclear (the repository lists no data terms); the status does not depend on the training data.
Licence evidence: https://huggingface.co/tomtomtommi/LiteAnyStereoV2/blob/17788d91618646fa781a14462e2926a034b9f49d/README.md#L2
Input size
any H x W >= 32, padded to a multiple of 32
Tested up to (portable, fp16, RTX 5060 Ti 16 GB): las2-h 3000x4000 (8.5 GiB peak allocated, 11.7 reserved); las2-l 3000x4000 (7.0 GiB peak allocated, 7.4 reserved); las2-m 3000x4000 (7.0 GiB peak allocated, 7.4 reserved); las2-s 3000x4000 (7.0 GiB peak allocated, 7.4 reserved)
Speed and memory
Stereo, Middlebury 2014 half, 994x1482
| model | time (ms): original fp32 | original | portable | kernels | kernels speed-up: vs original | vs portable | peak memory (MiB, allocated / reserved): original | portable | kernels |
|---|---|---|---|---|---|---|---|---|---|
las2-h |
204.8 | 134.9 | 122.8 | 53.27 | 2.5x | 2.3x | 1022 / 1084 | 1050 / 1110 | 1021 / 1388 |
las2-l |
153.5 | 90.8 | 80.32 | 29.92 | 3.0x | 2.7x | 1010 / 1082 | 1036 / 1098 | 397 / 600 |
las2-m |
79.1 | 52.4 | 54.13 | 18.29 | 2.9x | 3.0x | 948 / 1000 | 972 / 1026 | 329 / 572 |
las2-s |
62.0 | 43.5 | 46.86 | 14.35 | 3.0x | 3.3x | 933 / 978 | 955 / 988 | 320 / 558 |
Startup and first call
| model | load (s): original | portable | kernels | first call (s): original | portable | kernels |
|---|---|---|---|---|---|---|
las2-h |
0.26 (375x1242) | 0.18 | 0.13† | 1.21 (375x1242) | 0.14† | 0.06† |
las2-l |
0.39 (375x1242) | 0.23 | 0.18† | 0.04 (375x1242) | 0.08† | 0.03† |
las2-m |
0.17 (375x1242) | 0.12 | 0.10† | 0.02 (375x1242) | 0.05† | 0.02† |
las2-s |
0.11 (375x1242) | 0.09 | 0.07† | 0.02 (375x1242) | 0.05† | 0.01† |
Fixed-size speed and peak memory, all models
| model | mode | 640x480 | 1280x720 | 1920x1080 | 12 MP 4000x3000 |
|---|---|---|---|---|---|
las2-h |
portable | 29.8 ms / 269 / 328 | 69.1 ms / 668 / 904 | 174.8 ms / 1386 / 2168 | 1408.7 ms / 8708 / 12018 |
las2-h |
kernels | 10.2 ms / 218 / 596 | 28.5 ms / 604 / 1636 | 73.6 ms / 1344 / 1648 | 512.5 ms / 9695 / 12606 |
las2-l |
portable | 25.5 ms / 286 / 328 | 51.1 ms / 653 / 700 | 110.1 ms / 1317 / 1404 | 824.0 ms / 7201 / 7598 |
las2-l |
kernels | 6.6 ms / 129 / 270 | 16.8 ms / 241 / 676 | 38.6 ms / 422 / 588 | 254.6 ms / 2174 / 2896 |
las2-m |
portable | 18.9 ms / 219 / 252 | 35.4 ms / 586 / 648 | 73.9 ms / 1249 / 1358 | 602.8 ms / 7135 / 7584 |
las2-m |
kernels | 4.0 ms / 96 / 258 | 9.2 ms / 187 / 618 | 21.5 ms / 346 / 516 | 146.5 ms / 1870 / 2684 |
las2-s |
portable | 16.1 ms / 202 / 234 | 31.0 ms / 569 / 632 | 64.4 ms / 1232 / 1342 | 532.1 ms / 7119 / 7566 |
las2-s |
kernels | 3.0 ms / 88 / 246 | 6.9 ms / 179 / 602 | 16.6 ms / 338 / 500 | 109.3 ms / 1859 / 2598 |
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 authors' code 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
Original = the authors' code, same checkpoint and protocol.
| variant | dataset | original | portable | kernels | note |
|---|---|---|---|---|---|
las2-s |
ETH3D two-view train, bad > 1 px % / EPE | 4.10 / 0.382 | 4.09 / 0.382 | possibly in the training data | |
las2-s |
Middlebury 2014 trainingH, bad > 2 px % / EPE | 16.18 / 3.460 | 16.18 / 3.462 | possibly in the training data | |
las2-m |
ETH3D two-view train, bad > 1 px % / EPE | 2.97 / 0.310 | 2.97 / 0.310 | 2.98 / 0.310 | possibly in the training data |
las2-m |
Middlebury 2014 trainingH, bad > 2 px % / EPE | 11.24 / 2.324 | 11.23 / 2.324 | 11.22 / 2.323 | possibly in the training data |
las2-l |
ETH3D two-view train, bad > 1 px % / EPE | 2.24 / 0.268 | 2.25 / 0.269 | possibly in the training data | |
las2-l |
Middlebury 2014 trainingH, bad > 2 px % / EPE | 11.85 / 2.614 | 11.85 / 2.614 | possibly in the training data | |
las2-h |
ETH3D two-view train, bad > 1 px % / EPE | 2.14 / 0.258 | 2.13 / 0.258 | 2.14 / 0.258 | possibly in the training data |
las2-h |
Middlebury 2014 trainingH, bad > 2 px % / EPE | 9.27 / 2.141 | 9.27 / 2.141 | 9.28 / 2.141 | possibly in the training data |
Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-10-27.
Parity with the original code
Every variant is checked against golden outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.05 px and p99 <= 0.5 px.
Limitations
- The original code's LAS1 (LiteAnyStereo.pth, another architecture) is not included.
- Portable LAS2-H is faster than the original code but allocates more memory.
Usage
from fastforks import stereo
model = stereo.load("las2-h") # portable PyTorch on the GPU if there is one, else the CPU
result = model("left.png", "right.png")
result.disparity # [H, W] float32 tensor, input resolution
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: las2-h, las2-l, las2-m, las2-s.
Kernels backend
stereo.load("las2-h", mode="kernels") runs our CUDA kernels (igev, cutlass_conv, cudnn_hacc, stereo_ops). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels 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.