MonSter (fastforks)

MonSter (Cheng et al. 2025): Depth Anything V2 (ViT-L) monocular branch + IGEV-style geometry volume, with stereo and mono ConvGRUs that refine each other, 32 iterations. The original README flags its monster-sceneflow checkpoint as buggy.

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/Junda24/MonSter at 49016c9a72b4.

Author weight repos: https://huggingface.co/cjd24/MonSter.

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
monster-mix-all@1	mix_all.pth	fe0f25eacc4d	LicenseRef-noncommercial AND CC-BY-NC-4.0	no
monster-sceneflow@1	sceneflow.pth	ece58ae6f8ff	LicenseRef-noncommercial AND CC-BY-NC-4.0	no

Licence

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

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

Licence evidence: https://github.com/Junda24/MonSter/blob/49016c9a72b4260c0d119c77bb14e78a132cea8c/README.md#L93

Input size

any H x W >= 32, padded to a multiple of 32 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): monster-mix-all 1988x2880 (6.8 GiB peak allocated, 9.0 reserved); monster-sceneflow 1988x2880 (6.8 GiB peak allocated, 9.0 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
monster-mix-all	4275	4277	1585	668	6.4x	2.4x	6382 / 6430	2336 / 2494	2356 / 2714

Startup and first call

model	load (s): original	portable	kernels	first call (s): original	portable	kernels
monster-mix-all	5.85 (375x1242)	4.83	0.95†	1.26 (375x1242)	1.59†	0.67†

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)
monster-sceneflow	994x1482	1585	1096.3	102.6	671.0	384	668	2336 / 2494	2081 / 2272	2018

Fixed-size speed and peak memory, all models

model	mode	480x640	720x1280	1080x1920	12 MP 3000x4000
monster-mix-all	portable	284.1 ms / 1089 / 1184	894.5 ms / 1732 / 2154	2291.3 ms (n=3) / 2935 / 3668	OOM
monster-mix-all	kernels	127.2 ms / 1112 / 1250	379.5 ms / 1752 / 2288	956.8 ms / 2939 / 3460	OOM
monster-sceneflow	portable	284.1 ms / 1089 / 1184	894.5 ms / 1732 / 2154	2290.9 ms (n=3) / 2935 / 3668	OOM
monster-sceneflow	kernels	127.1 ms / 1112 / 1250	379.4 ms / 1752 / 2288	956.9 ms / 2939 / 3460	OOM

Apple silicon (MPS)

model	portable	kernels	speed-up
monster-sceneflow	3344 ms / 3491 MiB (float16)	2756 ms / 4093 MiB	1.21x

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

• monster-sceneflow: same network and settings as monster-mix-all, measured once (rows not repeated).

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 fp32	original fp16	portable	kernels	note
monster-sceneflow	ETH3D two-view train, bad > 1 px % / EPE	1.34 / 0.222	1.33 / 0.221	1.34 / 0.223	1.34 / 0.220
monster-sceneflow	Middlebury 2014 trainingH, bad > 2 px % / EPE	10.20 / 1.922	10.20 / 1.922	10.18 / 1.921	10.16 / 1.920
monster-mix-all	ETH3D two-view train, bad > 1 px % / EPE	1.25 / 0.186	1.26 / 0.186	1.24 / 0.186	1.25 / 0.186	trained on this dataset's public images; not a zero-shot result
monster-mix-all	Middlebury 2014 trainingH, bad > 2 px % / EPE	7.82 / 2.402	7.82 / 2.398	7.85 / 2.395	7.84 / 2.402	trained on this dataset's public images; not a zero-shot result

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-05.

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 px and p99 <= 0.25 px.

Usage

from fastforks import stereo

# non-commercial weights: run with FASTFORKS_NONCOMMERCIAL=1 in the environment
model = stereo.load("monster-mix-all")  # 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: monster-mix-all, monster-sceneflow.

Kernels backend

stereo.load("monster-mix-all", mode="kernels") runs our CUDA kernels (igev, cutlass_conv, cudnn_hacc, fastbb). 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.