8000 fastforks · GitHub
[go: up one dir, main page]

Skip to content
8000
@fastforks

fastforks

Research forks that speed up open-source vision models: plain-PyTorch 'fast' branches and custom-kernel branches, with measured speed and memory.

fastforks

This Organization holds many separate forks. It's purpose was superseded by the package: https://github.com/kleinicke/fastforks that contains the same repos and more in an easily to install and use form.

Research forks that make open-source vision models faster and lighter, without changing their results. Everything is measured on one consumer GPU (RTX 5060 Ti, 16 GB) and reported as speed and memory, plus startup, first-call and new-input-size cost.

Branches: main/master = untouched upstream mirror; portable = plain PyTorch only (exports to ONNX / TensorRT, no build step); cuda-kernels = on top of portable, custom CUDA kernels for maximum speed. The baseline for every speed-up is upstream + simple flags (driver-side settings only, such as fp16/bf16 autocast, channels_last, cudnn.benchmark, TF32); upstream without flags is shown separately. Bundle repos such as fast-vision-models have no tier branches: the tier is a function argument.

Every model below, at its main setting (stereo: Middlebury-H 1482x994; fast-vision-models: the largest setting). Speed-ups are relative to upstream + flags; memory = peak allocated, cuda-kernels / flags; "-" = tier does not exist; "n/m" = not measured; "x (vs upstream)" = repo has no flags row. Each section links to its full table in OVERVIEW.md (all tiers in ms, peak MB alloc / reserved, startup / first call / new size, quality, branches, commit SHAs, all settings).

Contents

Stereo matching

Full table: Stereo matching in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, metric in cell (flags -> cuda-kernels)
AANet aanet Apache-2.0, yes 151.1 -> 108.5 (Midd-H) - 1.39x 0.56x ETH3D EPE [px]: 0.545 -> 0.547; Midd-H EPE [px]: 15.721 -> 15.724
AANet+ aanet Apache-2.0, yes 189.6 -> 129.9 (Midd-H) - 1.46x 0.63x ETH3D EPE [px]: 0.745 -> 0.755; Midd-H EPE [px]: 15.790 -> 15.838
BridgeDepth (rvc_pretrain) BridgeDepth MIT, yes 441.1 -> 435.8 (Midd-H) 0.99x 1.01x 1.63x ETH3D EPE [px]: 0.088 -> 0.088; Midd-H EPE [px]: 1.238 -> 1.237
Cas-GwcNet-c cascade-stereo MIT code, yes 353.1 -> 227.9 (Midd-H) 0.91x 1.55x 0.71x ETH3D EPE [px]: 0.512 -> 0.513; Midd-H EPE [px]: 8.496 -> 8.572
Cas-GwcNet-gc cascade-stereo MIT code, yes 822.2 -> 377.1 (Midd-H) 1.47x 2.18x 0.23x ETH3D EPE [px]: 0.820 -> 0.824; Midd-H EPE [px]: 6.541 -> 6.541
CFNet CFNet MIT, yes 633 -> 396.8 (Midd-H) - 1.60x 0.26x ETH3D EPE [px]: 0.399 -> 0.397; Midd-H EPE [px]: 15.959 -> 15.952
CoEx coex GPL-3.0, custom 49.7 -> 40.2 (Midd-H) - 1.24x 0.43x ETH3D EPE [px]: 18.523 -> 18.497; Midd-H EPE [px]: 17.603 -> 17.575
CREStereo (ETH3D ckpt, 20 it) CREStereo-Pytorch none, no 1382 -> 238 (Midd-H) 1.51x 5.81x 0.39x ETH3D EPE [px]: 0.185 -> 0.185; Midd-H EPE [px]: 0.755 -> 0.745
CroCo-Stereo croco CC BY-NC-SA, no 3992.4 -> 2791.1 portable (Midd-H) 1.43x - - ETH3D EPE [px]: 0.026 -> 0.026 (port.)
DEFOM-Stereo DEFOM-Stereo MIT, yes 3715.4 -> 655.1 (Midd-H) - 5.67x 0.23x ETH3D EPE [px]: 0.173 -> 0.172; Midd-H EPE [px]: 1.462 -> 1.456
DLNR High-frequency-Stereo-Matching-Network MIT, yes 910.9 -> 264.9 (Midd-H) - 3.44x 0.23x ETH3D EPE [px]: 2.624 -> 2.639; Midd-H EPE [px]: 3.115 -> 3.119
Fast-ACVNet Fast-ACVNet MIT, yes 75.3 -> 57.4 (Midd-H) - 1.31x 0.26x ETH3D EPE [px]: 0.926 -> 0.935; Midd-H EPE [px]: 9.545 -> 9.547
Fast-FoundationStereo (23-36-37, 8 it) Fast-FoundationStereo custom, no 318.7 -> 144 (Midd-H) 0.97x 2.21x 0.69x ETH3D EPE [px]: 0.324 -> 0.323; Midd-H EPE [px]: 0.314 -> 0.315
FoundationStereo (ViT-L, 32 it) FoundationStereo custom, no 3642 -> 1495 (Midd-H) 1.22x 2.44x 0.37x ETH3D EPE [px]: 0.251 -> 0.250; Midd-H EPE [px]: 0.242 -> 0.241
GwcNet GwcNet MIT, yes 335.5 -> 279.9 (Midd-H) - 1.20x 0.30x ETH3D EPE [px]: 1.579 -> 1.579; Midd-H EPE [px]: 16.619 -> 16.623
IGEV-Stereo (sceneflow, 32 it) IGEV MIT code, yes 944.1 -> 293.3 (Midd-H) 1.22x 3.22x 0.69x ETH3D EPE [px]: 0.333 -> 0.333; Midd-H EPE [px]: 2.635 -> 2.636
LAS2-H LiteAnyStereo MIT, yes 118.7 -> 96.1 (Midd-H) - 1.24x 0.98x ETH3D EPE [px]: 0.154 -> 0.155; Midd-H EPE [px]: 6.142 -> 6.146
LAS2-M LiteAnyStereo MIT, yes 47.7 -> 40.6 (Midd-H) - 1.17x 0.72x ETH3D EPE [px]: 0.233 -> 0.233; Midd-H EPE [px]: 6.409 -> 6.409
MC-Stereo (Scene Flow ckpt, 32 it) MC-Stereo none, no n/m - - - EPE [px] vs GT (ETH3D -> Midd-H subset): 4.56 -> 1.14 -> -
MoCha-Stereo (V1, 32 it) MoCha-Stereo MIT code, yes 1395 -> 487 (1482x994) 1.17x 2.86x 0.48x EPE [px] ETH3D -> Midd-H: 0.318 -> 0.938 vs 0.318 -> 0.939
MonSter MonSter custom, custom 3152.3 -> 654.9 (Midd-H) - 4.81x 0.22x ETH3D EPE [px]: 0.215 -> 0.220; Midd-H EPE [px]: 5.255 -> 5.251
MonSter++ MonSter-plusplus custom, custom 3194.3 -> 676.2 (Midd-H) - 4.72x 0.23x ETH3D EPE [px]: 0.106 -> 0.106; Midd-H EPE [px]: 0.934 -> 0.933
NMRF-Stereo NMRF MIT, yes 416.7 -> 205.5 (Midd-H) - 2.03x 0.09x ETH3D EPE [px]: 0.338 -> 0.337; Midd-H EPE [px]: 6.961 -> 6.926
PromptStereo PromptStereo MIT, yes 3251.9 -> 1620 (Midd-H) - 2.01x 0.24x ETH3D EPE [px]: 0.116 -> 0.116; Midd-H EPE [px]: 1.087 -> 1.079
PSMNet PSMNet MIT, yes 526.5 -> 229 (Midd-H) - 2.30x 0.35x ETH3D EPE [px]: 1.255 -> 1.255; Midd-H EPE [px]: 24.939 -> 24.940
RAFT-Stereo (Middlebury ckpt, 32 it) RAFT-Stereo MIT code, yes 1018 -> 274 (Midd-H) 1.37x 3.72x 0.28x ETH3D EPE [px]: 0.288 -> 0.274; Midd-H EPE [px]: 1.069 -> 1.067
RT-MonSter++ MonSter-plusplus custom, custom 176.1 -> 119.5 (Midd-H) - 1.47x 0.42x ETH3D EPE [px]: 0.146 -> 0.147; Midd-H EPE [px]: 5.690 -> 5.685
S2M2-S s2m2 CC BY-NC, no 263 -> 111.4 (Midd-H) 0.75x 2.36x 0.62x ETH3D EPE [px]: 0.316 -> 0.193; Midd-H EPE [px]: 0.833 -> 0.652
S2M2-XL s2m2 CC BY-NC, no 1457 -> 683.8 (Midd-H) 0.80x 2.13x 1.13x ETH3D EPE [px]: 0.189 -> 0.149; Midd-H EPE [px]: 0.553 -> 0.499
Selective-Stereo Selective-Stereo MIT, yes 2148.7 -> 345.9 (Midd-H) - 6.21x 0.10x ETH3D EPE [px]: 0.403 -> 0.402; Midd-H EPE [px]: 7.503 -> 7.449
StereoAnywhere stereoanywhere CC BY-NC-SA, no 1801.9 -> 343.9 (KITTI 1242x375) - 5.24x 0.31x ETH3D EPE [px]: 0.143 -> 0.143; Midd-H EPE [px]: n/m -> n/m
UniMatch (stereo) unimatch MIT, yes 792.2 -> 517.1 portable (Midd-H) 1.53x - - ETH3D EPE [px]: 0.441 -> 0.441 (port.)
WAFT-Stereo B WAFT-Stereo custom, yes 1634.5 -> 410.2 (Midd-H) - 3.98x 0.17x ETH3D EPE [px]: 0.107 -> 0.108; Midd-H EPE [px]: 1.366 -> 1.365
WAFT-Stereo S WAFT-Stereo custom, yes 1129.3 -> 288.6 (Midd-H) - 3.91x 0.10x ETH3D EPE [px]: 0.096 -> 0.096; Midd-H EPE [px]: 1.744 -> 1.741

Multi-view stereo and feed-forward reconstruction

Full table: Multi-view stereo and feed-forward reconstruction in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, metric in cell (flags -> cuda-kernels)
MapAnything map-anything Apache-2.0, yes 1786 -> 556 (9 views) - 3.21x 0.31x ETH3D metric absrel: 0.1176 -> 0.1176
MVSAnywhere mvsanywhere NC, no 193.8 -> 194.7 portable (448x672) 1.00x - - ETH3D AbsRel: 0.0110 -> 0.0111 (port.)
MVSFormer++ MVSFormerPlusPlus Apache-2.0, yes 463.6 -> 273.7 (1536x1024) - 1.69x 0.27x n/m (random weights, accuracy not measured)
PatchmatchNet PatchmatchNet MIT, yes 313.3 -> 42.1 (1536x1024) - 7.44x 0.41x absrel: 0.1256 -> 0.1246
VGGT vggt NC, custom 952.1 -> 458.1 (8 frames) - 2.08x 0.28x flags identical to upstream; cuda depth rel. diff median 0.40% (8 frames)
VGGT-Omega vggt-omega NC, no 1001.6 -> 510 (size A) - 1.96x 0.25x ETH3D absrel: 0.0062 -> 0.0062

Monocular depth and geometry

Full table: Monocular depth and geometry in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, metric in cell (flags -> cuda-kernels)
Depth Pro ml-depth-pro custom, custom 528.1 -> 282.3 (Midd-H) 1.02x 1.87x 0.95x Midd-H AbsRel raw: 0.1480 -> 0.1481
Depth-Anything-3 (DA3MONO-LARGE) Depth-Anything-3 Apache-2.0 code, yes 175.4 -> 72.3 (1482x994 image) 1.50x 2.43x 0.47x NYU-20 AbsRel: 0.03394 -> 0.03397
Depth-Anything-V2 ViT-L Depth-Anything-V2 Apache-2.0 code, no 267.5 -> 75 (1024 px) 2.35x 3.57x 0.79x NYU AbsRel: 0.04388 -> 0.04388
Depth-Anything-V2 ViT-S Depth-Anything-V2 Apache-2.0 code, yes 66.2 -> 11.3 (1024 px) 3.96x 5.86x 0.79x NYU AbsRel: 0.05093 -> 0.05093
Metric3D v2 ViT-L Metric3D BSD-2, yes 438 -> 201.3 portable (Midd-H) 2.18x - - Midd-H AbsRel raw: 0.7041 -> 0.7040 (port.)
MiDaS dpt_beit_large_512 MiDaS MIT, yes 267 -> 84.2 portable (Midd-H) 3.17x - - Midd-H AbsRel (disparity): 0.1157 -> 0.1160 (port.)
MoGe-2 ViT-L normal MoGe MIT, yes 151.3 -> 85.9 (native 1482x994) 0.99x 1.76x 1.08x NYU-20 AbsRel: 0.0321 -> 0.0321
UniDepth V2 ViT-L UniDepth CC BY-NC, no 131.9 -> 62.7 (Midd-H) 1.31x 2.10x 0.45x Midd-H AbsRel raw: 0.1989 -> 0.1986
UniK3D ViT-L UniK3D CC BY-NC-SA, no 134.2 -> 62.6 (Midd-H) 1.33x 2.14x 0.44x Midd-H AbsRel raw: 0.2020 -> 0.2018

Promptable and instance segmentation

Full table: Promptable and instance segmentation in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, metric in cell (flags -> cuda-kernels)
Detectron2 (Mask R-CNN) detectron2 Apache-2.0, yes 40.8 -> 25.9 (size A) - 1.58x 0.68x box/mask AP: 45.38 -> 41.78 vs 45.18 -> 41.69
SAM 2 Hiera-L sam2 Apache-2.0, yes 86.4 -> 57.2 (image 1297x840) 1.22x 1.51x 0.70x IoU vs upstream (L): 1 -> 0.965
SAM 3 sam3 custom, custom 218 -> 115 (image 1297x840) 1.21x 1.90x 0.58x IoU vs GT: 0.7945 -> 0.7947
SAM ViT-B segment-anything Apache-2.0, yes 101.3 -> 41.5 (image 1297x840) 2.14x 2.44x 0.20x mean mask IoU vs fp32: 0.9906 -> 0.9902
SAM ViT-H segment-anything Apache-2.0, yes 308 -> 170.7 (image 1297x840) 1.44x 1.80x 0.47x mean mask IoU vs fp32: 0.9979 -> 0.9964
Sapiens-1B (seg) sapiens CC BY-NC, no 354.2 -> 189.3 (batch 1) - 1.87x 0.56x seg mIoU vs upstream: 0.9981 -> 0.9932

Semantic segmentation (fast-vision-models)

Full table: Semantic segmentation (fast-vision-models) in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, pixel agreement [%] (flags -> cuda-kernels)
ddrnet-23 fast-vision-models weights, yes 12.5 -> 7 (1024x2048) 1.02x 1.79x 0.55x 99.9 -> 99.9
deeplabv3-r101 fast-vision-models weights, yes 184 -> 95 (1024x2048) 1.01x 1.94x 0.58x 100.0 -> 100.0
deeplabv3plus-r101 fast-vision-models weights, yes 211 -> 101 (1024x2048) 1.17x 2.09x 0.46x 100.0 -> 100.0
dinov2-l-linear fast-vision-models weights, yes 497 -> 358 (1024x2048) 1.02x 1.39x 0.69x 100.0 -> 99.9
eomt-l-dinov2 fast-vision-models weights, custom 205 -> 142 (1024x2048) 1.09x 1.44x 0.54x 100.0 -> 100.0
eomt-l-dinov3 fast-vision-models weights, custom 299 -> 167 (1024x2048) 1.40x 1.79x 0.47x 100.0 -> 100.0
fcn-r101 fast-vision-models weights, yes 160 -> 82.1 (1024x2048) 1.01x 1.95x 0.50x 100.0 -> 100.0
hrnet-w48 fast-vision-models weights, yes 84.5 -> 52.3 (1024x2048) 1.02x 1.62x 0.27x 100.0 -> 100.0
mask2former-swinl fast-vision-models weights, no 279 -> 154 (1024x2048) 1.09x 1.81x 0.55x 99.9 -> 99.9
oneformer-swinl fast-vision-models weights, yes 335 -> 221 (1024x2048) 1.04x 1.52x 0.53x 99.8 -> 99.9
pidnet-l fast-vision-models weights, yes 20.7 -> 12.5 (1024x2048) 1.02x 1.66x 0.72x 99.9 -> 99.9
pspnet-r101 fast-vision-models weights, yes 191 -> 116 (1024x2048) 1.01x 1.65x 0.61x 100.0 -> 100.0
segformer-b0 fast-vision-models weights, no 34.3 -> 16.3 (1024x2048) 1.22x 2.10x 0.19x 100.0 -> 100.0
segformer-b5 fast-vision-models weights, no 135 -> 92.3 (1024x2048) 1.25x 1.46x 0.09x 100.0 -> 100.0
segnet fast-vision-models weights, n/a 144 -> 88.1 (1024x2048) 1.19x 1.63x 0.47x 97.2 -> 95.9
segnext-t fast-vision-models weights, yes 68.7 -> 26.4 (1024x2048) 1.76x 2.60x 0.19x 99.9 -> 99.9
unet-s5 fast-vision-models weights, yes 128 -> 73 (1024x2048) 1.02x 1.75x 0.63x 99.9 -> 99.8

Classification (fast-vision-models)

Full table: Classification (fast-vision-models) in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, top-1 agreement [%] (flags -> cuda-kernels)
alexnet fast-vision-models weights, custom 3.33 -> 1.99 (b64, per batch) 0.97x 1.67x 0.85x 99.6 -> 99.2
convnext_v2_b fast-vision-models weights, yes 107 -> 69 (b64, per batch) 1.38x 1.55x 1.02x 100.0 -> 100.0
deit_b fast-vision-models weights, yes 45.3 -> 38.9 (b64, per batch) 1.00x 1.16x 1.29x 100.0 -> 100.0
densenet121 fast-vision-models weights, custom 35.1 -> 16.5 (b64, per batch) 1.00x 2.13x 0.66x 100.0 -> 100.0
dinov2_l_lin fast-vision-models weights, yes 198 -> 168 (b64, per batch) 1.04x 1.18x 1.44x 100.0 -> 99.6
efficientnet_b0 fast-vision-models weights, yes 13.3 -> 7.74 (b64, per batch) 1.02x 1.72x 0.29x 100.0 -> 100.0
efficientnet_b4 fast-vision-models weights, yes 121 -> 66.9 (b64, per batch) 1.01x 1.81x 0.23x 100.0 -> 99.6
efficientnet_v2_s fast-vision-models weights, yes 91.1 -> 49 (b64, per batch) 1.01x 1.86x 0.50x 100.0 -> 100.0
efficientvit_b2 fast-vision-models weights, yes 55.5 -> 20.2 (b64, per batch) 1.01x 2.75x 0.21x 99.6 -> 100.0
eva02_l fast-vision-models weights, yes 1355 -> 853 (b64, per batch) 1.00x 1.59x 1.01x 99.6 -> 100.0
fastvit_sa24 fast-vision-models weights, custom 72.1 -> 22.82 (b64, per batch) 1.50x 3.16x 0.33x 100.0 -> 100.0
inception_v3 fast-vision-models weights, custom 45 -> 19.6 (b64, per batch) 1.01x 2.30x 0.51x 99.6 -> 99.6
mobilenet_v1 fast-vision-models weights, custom 9.07 -> 3.2 (b64, per batch) 1.01x 2.83x 0.29x 99.2 -> 100.0
mobilenet_v2 fast-vision-models weights, custom 10.9 -> 3.9 (b64, per batch) 1.03x 2.79x 0.21x 100.0 -> 98.4
mobilenet_v3_l fast-vision-models weights, custom 7.2 -> 3.76 (b64, per batch) 1.09x 1.91x 0.27x 99.6 -> 100.0
mobilenet_v4_m fast-vision-models weights, custom 12.3 -> 5.79 (b64, per batch) 1.04x 2.12x 0.34x 100.0 -> 99.6
regnety_16gf fast-vision-models weights, yes 102 -> 41 (b64, per batch) 1.01x 2.49x 0.39x 99.6 -> 99.2
resnet152 fast-vision-models weights, custom 66.7 -> 29.3 (b64, per batch) 1.00x 2.28x 0.70x 99.6 -> 99.6
resnet18 fast-vision-models weights, custom 9.62 -> 4.59 (b64, per batch) 1.00x 2.10x 0.49x 99.6 -> 100.0
resnet50 fast-vision-models weights, custom 30.9 -> 12.8 (b64, per batch) 1.00x 2.41x 0.66x 100.0 -> 100.0
shufflenet_v2 fast-vision-models weights, custom 5.57 -> 1.88 (b64, per batch) 1.01x 2.96x 0.80x 98.8 -> 99.6
siglip2_so400m_zs fast-vision-models weights, yes 807 -> 711 (b64, per batch) 1.00x 1.14x 0.89x 99.6 -> 99.2
siglip_so400m_ft fast-vision-models weights, yes 798 -> 711 (b64, per batch) 1.00x 1.12x 1.17x 99.6 -> 99.6
squeezenet1_1 fast-vision-models weights, custom 6.64 -> 1.95 (b64, per batch) 0.98x 3.41x 0.38x 99.6 -> 99.6
swin_b fast-vision-models weights, yes 89.9 -> 52.7 (b64, per batch) 1.25x 1.71x 0.91x 100.0 -> 100.0
vgg16 fast-vision-models weights, custom 67.2 -> 28.8 (b64, per batch) 1.00x 2.33x 0.43x 100.0 -> 99.6
vit_b16 fast-vision-models weights, yes 45.3 -> 38.8 (b64, per batch) 1.00x 1.17x 1.29x 100.0 -> 100.0

Backbones

Full table: Backbones in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, metric in cell (flags -> cuda-kernels)
DINOv2 ViT-L fastbb Apache-2.0, yes upstream 175.6 -> 114 (896x1344 b1) - 1.54x (vs upstream) - rel. L2 of patch tokens < 0.01 (cuda)
ResNet-50 fastbb Apache-2.0, yes upstream 9.27 -> 4.29 (1024x1024 b1) - 2.16x (vs upstream) - n/m
ViT-B/16 (timm) fastbb Apache-2.0, yes upstream 32.3 -> 21.6 (1024 px b1) - 1.50x (vs upstream) - n/m

Backbones (fast-vision-models)

Full table: Backbones (fast-vision-models) in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, feature rel. L2 error [ratio] (flags -> cuda-kernels)
dinov2-vitb14 fast-vision-models weights, yes 52.47 -> 39.75 (896x1344 b1) 1.01x 1.32x 1.39x 0.0077 -> 0.0116
dinov2-vitg14 fast-vision-models weights, yes 444.39 -> 364.43 (896x1344 b1) 1.01x 1.22x 1.09x 0.0036 -> 0.0037
dinov2-vitl14 fast-vision-models weights, yes 146.54 -> 113.64 (896x1344 b1) 1.01x 1.29x 1.13x 0.0066 -> 0.0073
dinov2-vits14 fast-vision-models weights, yes 23.73 -> 16.66 (896x1344 b1) 1.01x 1.42x 1.54x 0.0045 -> 0.0063
dinov3-convnext-base fast-vision-models weights, custom 25.62 -> 19.37 (1024px b1) 1.00x 1.32x 1.10x 0.0016 -> 0.0034
dinov3-convnext-large fast-vision-models weights, custom 49.62 -> 35.66 (1024px b1) 1.00x 1.39x 1.06x 0.0013 -> 0.0045
dinov3-convnext-small fast-vision-models weights, custom 17.54 -> 13.55 (1024px b1) 1.00x 1.29x 1.11x 0.0017 -> 0.0032
dinov3-convnext-tiny fast-vision-models weights, custom 10.9 -> 8.73 (1024px b1) 1.00x 1.25x 1.15x 0.0024 -> 0.0052
dinov3-vitb16 fast-vision-models weights, custom 32.31 -> 24.3 (1024px b1) 0.96x 1.33x 1.20x 0.0042 -> 0.0044
dinov3-vith16plus fast-vision-models weights, custom 207.14 -> 164.69 (1024px b1) 0.96x 1.26x 1.05x 0.0040 -> 0.0042
dinov3-vitl16 fast-vision-models weights, custom 112.9 -> 66.38 (1024px b1) 1.17x 1.70x 0.65x 0.0038 -> 0.0045
dinov3-vits16 fast-vision-models weights, custom 14.54 -> 10.09 (1024px b1) 0.95x 1.44x 1.50x 0.0034 -> 0.0032
timm-convnext_tiny.fb_in22k_ft_in1k fast-vision-models weights, yes 15.41 -> 13.33 (224px b32, per batch) 1.00x 1.16x 1.02x 0.0019, top-1 agree 1.000 -> 0.0027
timm-efficientnet_b0.ra_in1k fast-vision-models weights, yes 6.17 -> 4.7 (224px b32, per batch) 1.10x 1.31x 0.80x 0.0034, top-1 agree 1.000 -> 0.0041
timm-mobilenetv2_100.ra_in1k fast-vision-models weights, yes 4.83 -> 3.04 (224px b32, per batch) 1.14x 1.59x 0.79x 0.0053, top-1 agree 1.000 -> 0.0068
timm-regnety_032.ra_in1k fast-vision-models weights, yes 10.83 -> 31.61 (1024px b1) 1.02x 0.34x 1.46x 0.0020, top-1 agree 1.000 -> 0.0020
timm-resnet101.a1h_in1k fast-vision-models weights, yes 21.52 -> 10.52 (224px b32, per batch) 1.03x 2.05x 0.95x 0.0014, top-1 agree 1.000 -> 0.0050
timm-resnet18.a1_in1k fast-vision-models weights, yes 4.71 -> 2.42 (224px b32, per batch) 1.00x 1.95x 0.93x 0.0005, top-1 agree 1.000 -> 0.0006
timm-resnet50.a1_in1k fast-vision-models weights, yes 14.12 -> 6.58 (224px b32, per batch) 1.02x 2.15x 0.95x 0.0007, top-1 agree 1.000 -> 0.0013

Point tracking

Full table: Point tracking in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, metric in cell (flags -> cuda-kernels)
CoTracker co-tracker CC BY-NC, no 1157 -> 622.6 (48 frames) - 1.86x 0.43x median track deviation vs upstream (px): 0.0071 -> 0.0185

Gaussian splatting libraries

Full table: Gaussian splatting libraries in OVERVIEW.md.

model repo licence, commercial latency flags -> cuda-kernels [ms] (setting) portable speed-up vs flags cuda-kernels speed-up vs flags memory ratio (cuda-kernels / flags) quality, metric in cell (flags -> cuda-kernels)
fused-ssim fused-ssim MIT, yes upstream 1.39 -> 0.63 (L1+SSIM fwd+bwd) - 2.21x (vs upstream) - -
gsplat gsplat Apache-2.0, yes upstream 223.6 s -> 69.1 s (MipNeRF-360) - 3.24x (vs upstream) - PSNR [dB]: - -> 26.25 (upstream 26.24)

Notes

  • These are research forks. Upstream licences apply to the upstream code; our own additions in new repositories are Apache-2.0.
  • Much of the engineering was done with AI assistance (Claude); all results are measured, not estimated.
  • Kernels are built for sm_80 - sm_120 GPUs (Ampere to Blackwell); Jetson / Thor builds compile but are untested on hardware.

Maintained by Florian Nick.

Popular repositories Loading

  1. .github .github Public

    Organisation profile for fastforks

  2. fastbb fastbb Public

    Fast general drop-in backbones for PyTorch inference (DINOv2/ViT, CNN, RAFT-style encoders)

    Python

  3. s2m2 s2m2 Public

    Forked from junhong-3dv/s2m2

    Official implementation of "S²M²: Scalable Stereo Matching Model for Reliable Depth Estimation, ICCV 2025"

    Python

  4. RAFT-Stereo RAFT-Stereo Public

    Forked from princeton-vl/RAFT-Stereo

    Python

  5. CREStereo-Pytorch CREStereo-Pytorch Public

    Forked from ibaiGorordo/CREStereo-Pytorch

    Non-official Pytorch implementation of the CREStereo(CVPR 2022 Oral).

    Python

  6. gsplat gsplat Public

    Forked from nerfstudio-project/gsplat

    gsplat with fused CUDA training kernels (research fork of nerfstudio-project/gsplat)

    Python

Repositories

Showing 10 of 53 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…

0