Model card from fastforks 03e6fdbc

bd8a92e verified 33 minutes ago

12.1 kB

---
library_name: "fastforks"
pipeline_tag: "depth-estimation"
license: "other"
license_name: "mit-code-weights-unstated"
license_link: "https://github.com/Windsrain/Selective-Stereo/blob/c39acbfef9fbe36fbc01efa4abeafc62bbe2da36/README.md#L25"
tags:
- "stereo"
- "stereo-matching"
- "disparity"
- "cuda"
- "pytorch"
- "fastforks"
---
# Selective-IGEV (fastforks)

Selective-IGEV (Wang et al. 2024, Selective-Stereo): IGEV-Stereo with selective recurrent units, attention-blended small (1x1) and large (3x3) ConvGRUs on a channel- and spatially re-weighted context, 32 iterations.

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/Windsrain/Selective-Stereo> at `c39acbfef9fb`.

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`](https://github.com/kleinicke/fastforks/blob/main/src/fastforks/stereo/models/selective_igev/card.toml).

| variant | weights (author link) | SHA-256 | weights licence | commercial use |
|---|---|---|---|---|
| `selective-igev-eth3d-finetune@1` | [eth3d_finetune.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1fHtm5bno6usdcKEYLgH__yBLhMYDz0fs) | `29d62b91c58e` | MIT (assumed from code) | yes |
| `selective-igev-eth3d-train@1` | [eth3d_train.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1IQ6eP7oVy2MfRB9pIU0afPiXhnBT4QbG) | `989327ad0220` | MIT (assumed from code) | yes |
| `selective-igev-kitti12@1` | [kitti12.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1FS5LapV8uP19hce5pXw0wXMf7ZTz0djf) | `3a9ed6f2e778` | MIT (assumed from code) | yes |
| `selective-igev-kitti15@1` | [kitti15.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1ewtQTWvmPa4iGlRzIq4fcfk78LIq-Rim) | `488254b5cfbf` | MIT (assumed from code) | yes |
| `selective-igev-middlebury-finetune@1` | [middlebury_finetune.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1eBi5DC_mFfMNjvA13XhxpHkoP2ScUTqN) | `ce019e97bbe8` | MIT (assumed from code) | yes |
| `selective-igev-middlebury-train@1` | [middlebury_train.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1WDFD3qNaklB_OQDxHAdnMFNWCd678iAI) | `a51f67029d77` | MIT (assumed from code) | yes |
| `selective-igev-sceneflow@1` | [sceneflow.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1N0VfCFcAwSEPeXdWcrSS0JlNa4xm2gIn) | `27286b184d68` | MIT (assumed from code) | yes |

## Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: MIT. Files: `LICENSE`.

Code licence: MIT; the authors do not state a separate weights licence.

Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa).

Licence evidence: <https://github.com/Windsrain/Selective-Stereo/blob/c39acbfef9fbe36fbc01efa4abeafc62bbe2da36/README.md#L25>

## Input size

any H x W >= 32, padded to a multiple of 32
Tested up to (portable, fp16, RTX 5060 Ti 16 GB): `selective-igev-eth3d-finetune` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-eth3d-train` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-kitti12` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-kitti15` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-middlebury-finetune` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-middlebury-train` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-sceneflow` 3000x4000 (10.3 GiB peak allocated, 14.0 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 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| `selective-igev-eth3d-finetune` | 2201.3 | 1455.0 | 998.4 | 348.76 | 4.2x | 2.9x | 2013 / 2152 | 1238 / 1488 | 1109 / 1390 |

### Startup and first call

| model | load (s): original | portable | kernels | first call (s): original | portable | kernels |
| --- | --- | --- | --- | --- | --- | --- |
| `selective-igev-eth3d-finetune` | 0.36 (375x1242) | 0.50 | 0.14† | 2.25 (375x1242) | 1.02† | 0.35† |

### torch.compile and TensorRT

| model | input | portable ms | compile ms | compile s (cold) | TensorRT fp16 ms | ONNX + build s | kernels ms | peak MiB (alloc / reserved): portable | compile | TensorRT (device) |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| `selective-igev-eth3d-finetune` | 994x1482 | 998.4 | 690.7 | 125.8 | 460.8 | 256 | 348.76 | 1238 / 1488 | 919 / 1088 | 1148 |

### Fixed-size speed and peak memory, all models

| model | mode | 640x480 | 1280x720 | 1920x1080 | 12 MP 4000x3000 |
| --- | --- | --- | --- | --- | --- |
| `selective-igev-eth3d-finetune` | portable | 190.5 ms / 282 / 362 | 575.3 ms / 769 / 1036 | 1420.2 ms / 1690 / 2030 | 8918.7 ms (n=3) / 10549 / 14320 |
| `selective-igev-eth3d-finetune` | kernels | 82.5 ms / 260 / 368 | 212.8 ms / 693 / 1010 | 472.5 ms / 1502 / 2046 | 2806.7 ms (n=3) / 6928 / 8800 |
| `selective-igev-eth3d-train` | portable | 190.5 ms / 282 / 362 | 575.3 ms / 769 / 1036 | 1420.4 ms / 1690 / 2030 | 8919.8 ms (n=3) / 10549 / 14320 |
| `selective-igev-eth3d-train` | kernels | 82.4 ms / 260 / 368 | 212.8 ms / 693 / 1010 | 472.6 ms / 1502 / 2046 | 2807.0 ms (n=3) / 6928 / 8800 |
| `selective-igev-kitti12` | portable | 190.6 ms / 282 / 362 | 575.3 ms / 769 / 1036 | 1420.5 ms / 1690 / 2030 | 8917.7 ms (n=3) / 10549 / 14320 |
| `selective-igev-kitti12` | kernels | 82.3 ms / 260 / 368 | 212.9 ms / 693 / 1010 | 471.9 ms / 1502 / 2046 | 2803.4 ms (n=3) / 6928 / 8800 |
| `selective-igev-kitti15` | portable | 190.6 ms / 282 / 362 | 575.3 ms / 769 / 1036 | 1420.8 ms / 1690 / 2030 | 8918.8 ms (n=3) / 10549 / 14320 |
| `selective-igev-kitti15` | kernels | 82.5 ms / 260 / 368 | 212.8 ms / 693 / 1010 | 471.9 ms / 1502 / 2046 | 2803.9 ms (n=3) / 6928 / 8800 |
| `selective-igev-middlebury-finetune` | portable | 190.6 ms / 282 / 362 | 575.6 ms / 769 / 1036 | 1420.8 ms / 1690 / 2030 | 8923.1 ms (n=3) / 10549 / 14320 |
| `selective-igev-middlebury-finetune` | kernels | 82.3 ms / 260 / 368 | 212.6 ms / 693 / 1010 | 472.1 ms / 1502 / 2046 | 2804.4 ms (n=3) / 6928 / 8800 |
| `selective-igev-middlebury-train` | portable | 190.7 ms / 282 / 362 | 575.4 ms / 769 / 1036 | 1420.5 ms / 1690 / 2030 | 8921.4 ms (n=3) / 10549 / 14320 |
| `selective-igev-middlebury-train` | kernels | 82.5 ms / 260 / 368 | 213.1 ms / 693 / 1010 | 472.4 ms / 1502 / 2046 | 2804.8 ms (n=3) / 6928 / 8800 |
| `selective-igev-sceneflow` | portable | 190.6 ms / 282 / 362 | 575.4 ms / 769 / 1036 | 1420.3 ms / 1690 / 2030 | 8917.6 ms (n=3) / 10549 / 14320 |
| `selective-igev-sceneflow` | kernels | 82.5 ms / 260 / 368 | 213.5 ms / 693 / 1010 | 473.3 ms / 1502 / 2046 | 2808.4 ms (n=3) / 6928 / 8800 |

### Apple silicon (MPS)

| model | `portable` | `kernels` | speed-up |
| --- | --- | --- | --- |
| `selective-igev-sceneflow` | 2292 ms / 1727 MiB (float16) | 1878 ms / 1704 MiB | 1.22x |

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

- `selective-igev-eth3d-train`, `selective-igev-kitti12`, `selective-igev-kitti15`, `selective-igev-middlebury-finetune`, `selective-igev-middlebury-train`, `selective-igev-sceneflow`: same network and settings as `selective-igev-eth3d-finetune`, 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 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 |
|---|---|---|---|---|---|
| `selective-igev-sceneflow` | ETH3D two-view train, bad > 1 px % / EPE | 6.07 / 1.430 | 6.05 / 1.407 | 6.07 / 1.415 |  |
| `selective-igev-sceneflow` | Middlebury 2014 trainingH, bad > 2 px % / EPE | 11.78 / 2.584 | 11.80 / 2.587 | 11.79 / 2.585 |  |
| `selective-igev-kitti12` | ETH3D two-view train, bad > 1 px % / EPE |  | 24.07 / 3.647 | 24.07 / 3.660 |  |
| `selective-igev-kitti12` | Middlebury 2014 trainingH, bad > 2 px % / EPE |  | 20.08 / 4.845 | 20.10 / 4.862 |  |
| `selective-igev-kitti15` | ETH3D two-view train, bad > 1 px % / EPE |  | 24.26 / 2.442 | 24.26 / 2.447 |  |
| `selective-igev-kitti15` | Middlebury 2014 trainingH, bad > 2 px % / EPE |  | 19.65 / 4.956 | 19.64 / 4.942 |  |
| `selective-igev-middlebury-finetune` | ETH3D two-view train, bad > 1 px % / EPE |  | 24.26 / 0.995 | 24.24 / 0.995 |  |
| `selective-igev-middlebury-finetune` | Middlebury 2014 trainingH, bad > 2 px % / EPE |  | 5.66 / 1.208 | 5.66 / 1.192 | trained on this dataset's public images; not a zero-shot result |
| `selective-igev-middlebury-train` | ETH3D two-view train, bad > 1 px % / EPE |  | 10.24 / 0.433 | 10.26 / 0.433 |  |
| `selective-igev-middlebury-train` | Middlebury 2014 trainingH, bad > 2 px % / EPE |  | 5.95 / 1.434 | 5.99 / 1.436 | trained on this dataset's public images; not a zero-shot result |
| `selective-igev-eth3d-finetune` | ETH3D two-view train, bad > 1 px % / EPE |  | 0.32 / 0.082 | 0.32 / 0.082 | trained on this dataset's public images; not a zero-shot result |
| `selective-igev-eth3d-finetune` | Middlebury 2014 trainingH, bad > 2 px % / EPE |  | 12.33 / 2.811 | 12.35 / 2.811 |  |
| `selective-igev-eth3d-train` | ETH3D two-view train, bad > 1 px % / EPE |  | 0.78 / 0.137 | 0.78 / 0.137 | trained on this dataset's public images; not a zero-shot result |
| `selective-igev-eth3d-train` | Middlebury 2014 trainingH, bad > 2 px % / EPE |  | 7.99 / 2.162 | 8.01 / 2.156 |  |
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| <= 0.02 px and p99 <= 0.25 px.
## Limitations
- ONNX export needs torch >= 2.14 (older exporters fail on this network).
## Usage
```python
from fastforks import stereo
model = stereo.load("selective-igev-eth3d-finetune")  # mode="auto": the preferred backend from recorded measurements
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: `selective-igev-eth3d-finetune`, `selective-igev-eth3d-train`, `selective-igev-kitti12`, `selective-igev-kitti15`, `selective-igev-middlebury-finetune`, `selective-igev-middlebury-train`, `selective-igev-sceneflow`.
## Kernels backend
`stereo.load("selective-igev-eth3d-finetune", mode="kernels")` runs our CUDA kernels (`igev`, `cutlass_conv`, `cudnn_hacc`). 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.
<!-- staged from fastforks 03e6fdbc (2026-11-05) -->

	---
	library_name: "fastforks"
	pipeline_tag: "depth-estimation"
	license: "other"
	license_name: "mit-code-weights-unstated"
	license_link: "https://github.com/Windsrain/Selective-Stereo/blob/c39acbfef9fbe36fbc01efa4abeafc62bbe2da36/README.md#L25"
	tags:
	- "stereo"
	- "stereo-matching"
	- "disparity"
	- "cuda"
	- "pytorch"
	- "fastforks"
	---
	# Selective-IGEV (fastforks)

	Selective-IGEV (Wang et al. 2024, Selective-Stereo): IGEV-Stereo with selective recurrent units, attention-blended small (1x1) and large (3x3) ConvGRUs on a channel- and spatially re-weighted context, 32 iterations.

	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/Windsrain/Selective-Stereo> at `c39acbfef9fb`.

	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`](https://github.com/kleinicke/fastforks/blob/main/src/fastforks/stereo/models/selective_igev/card.toml).

	\| variant \| weights (author link) \| SHA-256 \| weights licence \| commercial use \|
	\|---\|---\|---\|---\|---\|
	\| `selective-igev-eth3d-finetune@1` \| [eth3d_finetune.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1fHtm5bno6usdcKEYLgH__yBLhMYDz0fs) \| `29d62b91c58e` \| MIT (assumed from code) \| yes \|
	\| `selective-igev-eth3d-train@1` \| [eth3d_train.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1IQ6eP7oVy2MfRB9pIU0afPiXhnBT4QbG) \| `989327ad0220` \| MIT (assumed from code) \| yes \|
	\| `selective-igev-kitti12@1` \| [kitti12.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1FS5LapV8uP19hce5pXw0wXMf7ZTz0djf) \| `3a9ed6f2e778` \| MIT (assumed from code) \| yes \|
	\| `selective-igev-kitti15@1` \| [kitti15.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1ewtQTWvmPa4iGlRzIq4fcfk78LIq-Rim) \| `488254b5cfbf` \| MIT (assumed from code) \| yes \|
	\| `selective-igev-middlebury-finetune@1` \| [middlebury_finetune.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1eBi5DC_mFfMNjvA13XhxpHkoP2ScUTqN) \| `ce019e97bbe8` \| MIT (assumed from code) \| yes \|
	\| `selective-igev-middlebury-train@1` \| [middlebury_train.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1WDFD3qNaklB_OQDxHAdnMFNWCd678iAI) \| `a51f67029d77` \| MIT (assumed from code) \| yes \|
	\| `selective-igev-sceneflow@1` \| [sceneflow.pth](https://drive.usercontent.google.com/download?export=download&confirm=t&id=1N0VfCFcAwSEPeXdWcrSS0JlNa4xm2gIn) \| `27286b184d68` \| MIT (assumed from code) \| yes \|

	## Licence

	Original-code strings, recorded as they are (research-use package; check them before any other use). Code: MIT. Files: `LICENSE`.

	Code licence: MIT; the authors do not state a separate weights licence.

	Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa).

	Licence evidence: <https://github.com/Windsrain/Selective-Stereo/blob/c39acbfef9fbe36fbc01efa4abeafc62bbe2da36/README.md#L25>

	## Input size

	any H x W >= 32, padded to a multiple of 32
	Tested up to (portable, fp16, RTX 5060 Ti 16 GB): `selective-igev-eth3d-finetune` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-eth3d-train` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-kitti12` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-kitti15` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-middlebury-finetune` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-middlebury-train` 3000x4000 (10.3 GiB peak allocated, 14.0 reserved); `selective-igev-sceneflow` 3000x4000 (10.3 GiB peak allocated, 14.0 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 \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| `selective-igev-eth3d-finetune` \| 2201.3 \| 1455.0 \| 998.4 \| 348.76 \| 4.2x \| 2.9x \| 2013 / 2152 \| 1238 / 1488 \| 1109 / 1390 \|

	### Startup and first call

	\| model \| load (s): original \| portable \| kernels \| first call (s): original \| portable \| kernels \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| `selective-igev-eth3d-finetune` \| 0.36 (375x1242) \| 0.50 \| 0.14† \| 2.25 (375x1242) \| 1.02† \| 0.35† \|

	### torch.compile and TensorRT

	\| model \| input \| portable ms \| compile ms \| compile s (cold) \| TensorRT fp16 ms \| ONNX + build s \| kernels ms \| peak MiB (alloc / reserved): portable \| compile \| TensorRT (device) \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| `selective-igev-eth3d-finetune` \| 994x1482 \| 998.4 \| 690.7 \| 125.8 \| 460.8 \| 256 \| 348.76 \| 1238 / 1488 \| 919 / 1088 \| 1148 \|

	### Fixed-size speed and peak memory, all models

	\| model \| mode \| 640x480 \| 1280x720 \| 1920x1080 \| 12 MP 4000x3000 \|
	\| --- \| --- \| --- \| --- \| --- \| --- \|
	\| `selective-igev-eth3d-finetune` \| portable \| 190.5 ms / 282 / 362 \| 575.3 ms / 769 / 1036 \| 1420.2 ms / 1690 / 2030 \| 8918.7 ms (n=3) / 10549 / 14320 \|
	\| `selective-igev-eth3d-finetune` \| kernels \| 82.5 ms / 260 / 368 \| 212.8 ms / 693 / 1010 \| 472.5 ms / 1502 / 2046 \| 2806.7 ms (n=3) / 6928 / 8800 \|
	\| `selective-igev-eth3d-train` \| portable \| 190.5 ms / 282 / 362 \| 575.3 ms / 769 / 1036 \| 1420.4 ms / 1690 / 2030 \| 8919.8 ms (n=3) / 10549 / 14320 \|
	\| `selective-igev-eth3d-train` \| kernels \| 82.4 ms / 260 / 368 \| 212.8 ms / 693 / 1010 \| 472.6 ms / 1502 / 2046 \| 2807.0 ms (n=3) / 6928 / 8800 \|
	\| `selective-igev-kitti12` \| portable \| 190.6 ms / 282 / 362 \| 575.3 ms / 769 / 1036 \| 1420.5 ms / 1690 / 2030 \| 8917.7 ms (n=3) / 10549 / 14320 \|
	\| `selective-igev-kitti12` \| kernels \| 82.3 ms / 260 / 368 \| 212.9 ms / 693 / 1010 \| 471.9 ms / 1502 / 2046 \| 2803.4 ms (n=3) / 6928 / 8800 \|
	\| `selective-igev-kitti15` \| portable \| 190.6 ms / 282 / 362 \| 575.3 ms / 769 / 1036 \| 1420.8 ms / 1690 / 2030 \| 8918.8 ms (n=3) / 10549 / 14320 \|
	\| `selective-igev-kitti15` \| kernels \| 82.5 ms / 260 / 368 \| 212.8 ms / 693 / 1010 \| 471.9 ms / 1502 / 2046 \| 2803.9 ms (n=3) / 6928 / 8800 \|
	\| `selective-igev-middlebury-finetune` \| portable \| 190.6 ms / 282 / 362 \| 575.6 ms / 769 / 1036 \| 1420.8 ms / 1690 / 2030 \| 8923.1 ms (n=3) / 10549 / 14320 \|
	\| `selective-igev-middlebury-finetune` \| kernels \| 82.3 ms / 260 / 368 \| 212.6 ms / 693 / 1010 \| 472.1 ms / 1502 / 2046 \| 2804.4 ms (n=3) / 6928 / 8800 \|
	\| `selective-igev-middlebury-train` \| portable \| 190.7 ms / 282 / 362 \| 575.4 ms / 769 / 1036 \| 1420.5 ms / 1690 / 2030 \| 8921.4 ms (n=3) / 10549 / 14320 \|
	\| `selective-igev-middlebury-train` \| kernels \| 82.5 ms / 260 / 368 \| 213.1 ms / 693 / 1010 \| 472.4 ms / 1502 / 2046 \| 2804.8 ms (n=3) / 6928 / 8800 \|
	\| `selective-igev-sceneflow` \| portable \| 190.6 ms / 282 / 362 \| 575.4 ms / 769 / 1036 \| 1420.3 ms / 1690 / 2030 \| 8917.6 ms (n=3) / 10549 / 14320 \|
	\| `selective-igev-sceneflow` \| kernels \| 82.5 ms / 260 / 368 \| 213.5 ms / 693 / 1010 \| 473.3 ms / 1502 / 2046 \| 2808.4 ms (n=3) / 6928 / 8800 \|

	### Apple silicon (MPS)

	\| model \| `portable` \| `kernels` \| speed-up \|
	\| --- \| --- \| --- \| --- \|
	\| `selective-igev-sceneflow` \| 2292 ms / 1727 MiB (float16) \| 1878 ms / 1704 MiB \| 1.22x \|

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

	- `selective-igev-eth3d-train`, `selective-igev-kitti12`, `selective-igev-kitti15`, `selective-igev-middlebury-finetune`, `selective-igev-middlebury-train`, `selective-igev-sceneflow`: same network and settings as `selective-igev-eth3d-finetune`, 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 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 \|
	\|---\|---\|---\|---\|---\|---\|
	\| `selective-igev-sceneflow` \| ETH3D two-view train, bad > 1 px % / EPE \| 6.07 / 1.430 \| 6.05 / 1.407 \| 6.07 / 1.415 \| \|
	\| `selective-igev-sceneflow` \| Middlebury 2014 trainingH, bad > 2 px % / EPE \| 11.78 / 2.584 \| 11.80 / 2.587 \| 11.79 / 2.585 \| \|
	\| `selective-igev-kitti12` \| ETH3D two-view train, bad > 1 px % / EPE \| \| 24.07 / 3.647 \| 24.07 / 3.660 \| \|
	\| `selective-igev-kitti12` \| Middlebury 2014 trainingH, bad > 2 px % / EPE \| \| 20.08 / 4.845 \| 20.10 / 4.862 \| \|
	\| `selective-igev-kitti15` \| ETH3D two-view train, bad > 1 px % / EPE \| \| 24.26 / 2.442 \| 24.26 / 2.447 \| \|
	\| `selective-igev-kitti15` \| Middlebury 2014 trainingH, bad > 2 px % / EPE \| \| 19.65 / 4.956 \| 19.64 / 4.942 \| \|
	\| `selective-igev-middlebury-finetune` \| ETH3D two-view train, bad > 1 px % / EPE \| \| 24.26 / 0.995 \| 24.24 / 0.995 \| \|
	\| `selective-igev-middlebury-finetune` \| Middlebury 2014 trainingH, bad > 2 px % / EPE \| \| 5.66 / 1.208 \| 5.66 / 1.192 \| trained on this dataset's public images; not a zero-shot result \|
	\| `selective-igev-middlebury-train` \| ETH3D two-view train, bad > 1 px % / EPE \| \| 10.24 / 0.433 \| 10.26 / 0.433 \| \|
	\| `selective-igev-middlebury-train` \| Middlebury 2014 trainingH, bad > 2 px % / EPE \| \| 5.95 / 1.434 \| 5.99 / 1.436 \| trained on this dataset's public images; not a zero-shot result \|
	\| `selective-igev-eth3d-finetune` \| ETH3D two-view train, bad > 1 px % / EPE \| \| 0.32 / 0.082 \| 0.32 / 0.082 \| trained on this dataset's public images; not a zero-shot result \|
	\| `selective-igev-eth3d-finetune` \| Middlebury 2014 trainingH, bad > 2 px % / EPE \| \| 12.33 / 2.811 \| 12.35 / 2.811 \| \|
	\| `selective-igev-eth3d-train` \| ETH3D two-view train, bad > 1 px % / EPE \| \| 0.78 / 0.137 \| 0.78 / 0.137 \| trained on this dataset's public images; not a zero-shot result \|
	\| `selective-igev-eth3d-train` \| Middlebury 2014 trainingH, bad > 2 px % / EPE \| \| 7.99 / 2.162 \| 8.01 / 2.156 \| \|

	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\| <= 0.02 px and p99 <= 0.25 px.

	## Limitations

	- ONNX export needs torch >= 2.14 (older exporters fail on this network).

	## Usage

	```python
	from fastforks import stereo

	model = stereo.load("selective-igev-eth3d-finetune") # mode="auto": the preferred backend from recorded measurements
	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: `selective-igev-eth3d-finetune`, `selective-igev-eth3d-train`, `selective-igev-kitti12`, `selective-igev-kitti15`, `selective-igev-middlebury-finetune`, `selective-igev-middlebury-train`, `selective-igev-sceneflow`.

	## Kernels backend

	`stereo.load("selective-igev-eth3d-finetune", mode="kernels")` runs our CUDA kernels (`igev`, `cutlass_conv`, `cudnn_hacc`). 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.

	<!-- staged from fastforks 03e6fdbc (2026-11-05) -->