zenmetrics

Reproducible CPU benchmarks and per-OS installation. zenmetrics is a pure-Rust collection of full-reference image-quality metrics, with one CLI that scores any (reference, distorted) pair with any of them.

Most metrics are ports of the authors' published method, checked against a reference implementation; the table below says which, and by how much, and DIVERGENCES.md explains each difference. Where a dataset or tool computed a metric with its own conventions (the scores published with the JPEG AIC2026 dataset, libvmaf), that variant is available too, so published numbers can be reproduced. Six metrics also run on the GPU (CUDA, Vulkan, Metal, DX12) through CubeCL. unsafe_code = "forbid" is the workspace default; some crates override it.

Interactive browser viewer · Checkpoint downloads

Quick start

Nothing here is on crates.io yet (every crate is publish = false). The workspace path-depends on sibling Imazen repos, so clone those next to it first, exactly as CI does:

git clone https://github.com/imazen/zenmetrics && cd zenmetrics
bash scripts/ci/clone-siblings.sh            # clones the pinned siblings into ../
cargo build --release -p zenmetrics-cli      # binary: target/release/zenmetrics

Score one pair:

zenmetrics score --metric ssim2 --reference ref.png --distorted out.jpg

Score several encodes of one reference with several metrics, decoding each image once:

zenmetrics compare --reference ref.png \
  --variant q60.jpg --variant q80.jpg --variant out.avif \
  --metric ssim2 --metric butteraugli --metric dssim --output tsv

Score a TSV of pairs (ref_path, dist_path columns plus a header row; extra columns pass through to the output TSV):

zenmetrics batch --metric cvvdp --display-model standard_fhd \
  --pairs pairs.tsv --output scores.tsv --jobs 8

zenmetrics list-metrics prints what your build enabled and which metrics need a GPU; zenmetrics list-formats prints the decoders.

Checkpoint binaries package the CLI, IQA comparison drivers and fleet executors with pinned build identities.

Metrics

Models & metrics explorer · Full cards, papers and terms

Every CLI name is listed below. FR/NR indicates reference requirements; NN marks neural models. Code and model terms, paper citations and full reference identities remain in the linked cards.

Rust/C is Rust time divided by canonical time (lower is faster), using a representative warm cohort, preferring one thread. Approximate local comparisons are descriptive, with their input size, thread budget and source pins in the comparison ledger. A dash means no comparable measurement. “Prior” means older recorded evidence; “gate” means a declared bound, not a measured pass. Δ is the measured difference without an assigned bound. Pass counts include retained follow-up cohorts; they are not counts of distinct inputs. ! flags any observed warm case above 1.08; R/O identifies RTEN/ORT CPU for neural ratios. Cold numerical misses are shown separately from the retained warm evidence; See the ledger for confirmation status.

Metric	Type	Canonical	Check	Rust/C
`afine`	FR/NN	original A-FINE / pin…	2/2 pass	≈0.95× R
`ahiq`	FR/NN	original AHIQ / pinne…	2/2 pass	≈0.55× R
`ahiq-pyiqa`	FR/NN	PyIQA 0.1.16 / native…	2/2 pass	≈0.45× O
`arniqa`	NR/NN	original ARNIQA / kad…	5/5 pass	≈1.15× R !
`brisque`	NR	PIQ 0.8.0	2 cold misses	≈0.12×
`butteraugli`	FR	libjxl 0.12.0 origina…	Δ 5e-05	≈0.41×
`butteraugli-gpu`	FR	libjxl 0.12.0 origina…	Δ 0.0083	≈0.022×
`butteraugli-original`	FR	original google/butte…	Δ 0	≈0.7×
`cambi`	NR	libvmaf 3.2.0 / named…	4 process errors	≈0.884× !
`cambi-aic2026`	NR	libvmaf 3.2.0 / named…	2 process errors	≈0.86×
`ciede2000`	FR	scikit-image CIEDE200…	Δ 9.2e-15	≈0.6×
`ckdn`	FR/NN	original CKDN / pinne…	2/2 pass	≈0.49× R
`cvvdp`	FR	pycvvdp 0.5.7 CPU / s…	1/1 pass	≈0.031×
`cvvdp-gpu`	FR	CUDA pycvvdp 0.5.7 /…	17/49 pass	≈0.12×
`cw-ssim`	FR	author CW-SSIM / orig…	Δ 6.7e-16	≈0.062×
`cw-ssim-pyiqa`	FR	PyIQA 0.1.16 / weight…	26/26 pass	≈0.11×
`deepdc`	FR/NN	DeepDC-PyTorch 0.2 /…	5/5 pass	≈1.21× R !
`dists`	FR/NN	DISTS-pytorch 0.1	3/3 pass	≈1.11× R !
`dmm`	FR/NN	original DMM / pinned…	partial; 2/2 pass	≈0.61× R
`dmm-pyiqa`	FR/NN	PyIQA 0.1.16 / native…	partial; 2/2 pass	≈0.52× O
`dreamsim`	FR/NN	DreamSim 0.2.1 / ense…	2/2 pass	≈0.921× R
`dssim`	FR	author dssim-core3.5.…	Δ 0	≈0.23×
`dssim-gpu`	FR	author dssim-core 3.5…	Δ 0.00038	≈0.037× !
`flip`	FR	NVIDIA FLIP 1.7 / f64…	Δ 3.6e-14	≈0.18× !
`flip-v17`	FR	NVIDIA FLIP 1.7	Δ 0	≈0.19× !
`fsim`	FR	original FeatureSIM /…	6/6 pass	≈0.22×
`fsim-k2`	FR	original FeatureSIM /…	6/6 pass	≈0.18×
`fsim-k3`	FR	original FeatureSIM /…	6/6 pass	≈0.16×
`fsim-k4`	FR	original FeatureSIM /…	6/6 pass	≈0.14×
`fsim-nodownscale`	FR	original FeatureSIM /…	6/6 pass	≈0.21×
`fsim-y`	FR	author FeatureSIM.m f…	3/3 pass	≈0.046×
`gmsd`	FR	Clunie libgmsd de646c…	2/2 pass	≈0.091×
`haarpsi`	FR	original HaarPSI / Oc…	6/6 pass	≈0.12×
`haarpsi-y`	FR	original HaarPSI / Oc…	6/6 pass	≈0.13×
`hdrvdp`	FR	official HDR-VDP 2.2.…	scope caveat	—
`hdrvdp-aic2026`	FR	official HDR-VDP 2.2.…	Δ 0.092	≈0.12×
`hdrvdp3`	FR	official HDR-VDP 3.0.…	scope caveat	—
`hdrvdp3-aic2026`	FR	official HDR-VDP 3.0.…	Δ 0.00029	≈0.18×
`hyperiqa`	NR/NN	original HyperIQA / f…	3/3 pass	≈0.74× R !
`iwssim`	FR	Python-IW-SSIM f9de37…	3 process errors	≈0.084×
`iwssim-gpu`	FR	CUDA Python-IW-SSIM f…	5 process errors	≈0.14×
`iwssim-piq`	FR	author IW-SSIM / PIQ…	2/2 pass	≈0.072×
`liqe`	NR/NN	original LIQE / pinne…	2/2 pass	≈0.088× R
`lpips-alex`	FR/NN	lpips 0.1.4	5/5 pass	≈1.16× R !
`lpips-squeeze`	FR/NN	lpips 0.1.4	2/2 pass	≈0.64× R
`lpips-vgg`	FR/NN	lpips 0.1.4	4/4 pass	≈1.08× R !
`mad`	FR	author HI/LO + paper…	Δ 0	≈0.12×
`maniqa`	NR/NN	original MANIQA / KonIQ	2/2 pass	≈0.903× R
`mdctpsnr`	FR	thorfdbg mDCTpsnr 9a2…	2/2 pass	≈0.804×
`mdsi`	FR	original MDSI / Octav…	6/6 pass	≈0.21×
`ms-gmsd`	FR	unrecorded	paper variant	—
`ms-gmsdc`	FR	unrecorded	paper variant	—
`ms-ssim-pyiqa`	FR	PyIQA 0.1.16 / named…	4/4 pass	≈0.013×
`ms-ssim-pyiqa-f32`	FR	PyIQA 0.1.16 / named…	3/3 pass	≈0.013×
`ms-ssim-pyiqa-f32-k2`	FR	PyIQA 0.1.16 / named…	4/4 pass	≈0.026×
`ms-ssim-pyiqa-f32-k3`	FR	PyIQA 0.1.16 / named…	4/4 pass	≈0.048×
`ms-ssim-pyiqa-f32-k4`	FR	PyIQA 0.1.16 / named…	6/6 pass	≈0.075×
`ms-ssim-pyiqa-f32-k5`	FR	PyIQA 0.1.16 / named…	6/6 pass	≈0.21×
`ms-ssim-pyiqa-f32-k6`	FR	PyIQA 0.1.16 / named…	6/6 pass	≈0.26×
`msssim`	FR	original msssim / Oct…	15/30 pass	≈0.01×
`msssim-libvmaf`	FR	libvmaf 3.2.0 / named…	4 process errors	≈0.36×
`musiq`	NR/NN	Google MUSIQ 3ef5211…	4/4 pass	≈0.58× R
`niqe`	NR	BasicSR 1.4.2 / resid…	4 cold misses	≈0.18×
`nlpd`	FR	NLPD_Pytorch / origin…	Δ 6.5e-07	≈0.051×
`nlpd-iqa`	FR	IQA_pytorch NLPD / qu…	Δ 7.1e-08	≈0.063×
`paq2piq`	NR/NN	original Paq2piq / 40…	5/5 pass	≈1.3× R !
`pieapp`	FR/NN	Official v0.1 public…	excluded	—
`pieapp-pyiqa`	FR/NN	pyiqa 0.1.16 create_m…	excluded	—
`piqe`	NR	pypiqe 1.2	170/170 pass	≈0.1×
`proposal-butteraugli`	FR	libjxl 0.12.0 origina…	Δ 1.6e-08	≈0.4×
`psnr`	FR	scikit-image 0.26 / d…	Δ 0	≈0.0014×
`psnr-cb`	FR	scikit-image 0.26 / f…	Δ 0	≈0.098×
`psnr-cr`	FR	scikit-image 0.26 / f…	Δ 0	≈0.12×
`psnr-y`	FR	scikit-image 0.26 / d…	Δ 3.1e-11	≈0.05×
`psnr-y-libvmaf`	FR	libvmaf 3.2.0 / named…	Δ 0	≈0.16×
`psnr-y-studio601`	FR	scikit-image 0.26 / d…	Δ 0	≈0.011×
`psnr-y601`	FR	scikit-image 0.26 / d…	Δ 2.4e-12	≈0.04×
`psnr-ycbcr611`	FR	NumPy / declared floa…	Δ 0	≈0.024×
`psnrhvs`	FR	author psnrhvsm.m / s…	1 process errors	≈0.00016×
`psnrhvs-daala`	FR	libvmaf 3.2.0 / origi…	28/28 pass	≈0.57×
`psnrhvs-y`	FR	author psnrhvsm.m / s…	1 process errors	≈0.00029×
`ssim`	FR	scikit-image SSIM / d…	Δ 6.7e-16	≈0.15×
`ssim-libvmaf`	FR	libvmaf 3.2.0 / named…	Δ 2e-06	≈0.54×
`ssim-pyiqa`	FR	PyIQA 0.1.16 / named…	4/4 pass	≈0.013×
`ssim-pyiqa-f32`	FR	PyIQA 0.1.16 / named…	4/6 pass	≈0.014×
`ssim-pyiqa-f32-k2`	FR	PyIQA 0.1.16 / named…	3/3 pass	≈0.025×
`ssim-pyiqa-f32-k3`	FR	PyIQA 0.1.16 / named…	3/3 pass	≈0.041×
`ssim-pyiqa-f32-k4`	FR	PyIQA 0.1.16 / named…	3/3 pass	≈0.067×
`ssim-pyiqa-f32-k5`	FR	PyIQA 0.1.16 / named…	3/3 pass	≈0.16×
`ssim-pyiqa-f32-k6`	FR	PyIQA 0.1.16 / named…	3/3 pass	≈0.28×
`ssim2`	FR	primary SSIMULACRA2 v2.1	Δ 0	≈0.58×
`ssim2-gpu`	FR	primary SSIMULACRA2 v…	Δ 6	≈0.062× !
`ssimc`	FR	scikit-image 0.26 / d…	Δ 1.2e-15	≈0.24×
`ssimc-pyiqa-f32`	FR	PyIQA 0.1.16 / named…	3/3 pass	≈0.013×
`ssimc-pyiqa-f32-k2`	FR	PyIQA 0.1.16 / named…	3/3 pass	≈0.023×
`ssimc-pyiqa-f32-k3`	FR	PyIQA 0.1.16 / named…	4/4 pass	≈0.039×
`ssimc-pyiqa-f32-k4`	FR	PyIQA 0.1.16 / named…	4/4 pass	≈0.065×
`ssimc-pyiqa-f32-k5`	FR	PyIQA 0.1.16 / named…	4/4 pass	≈0.096×
`ssimc-pyiqa-f32-k6`	FR	PyIQA 0.1.16 / named…	4/4 pass	≈0.54×
`ssimulacra1`	FR	original SSIMULACRA1…	Δ 3.8e-09	≈1.07× !
`stlpips-alex`	FR/NN	stlpips-pytorch 0.0.2	5/5 pass	≈1.06× R !
`stlpips-vgg`	FR/NN	stlpips-pytorch 0.0.2	2/2 pass	≈0.958× R
`topiq-fr`	FR/NN	PyIQA 0.1.16 / native…	2/2 pass	≈1.05× O
`topiq-fr-pipal`	FR/NN	PyIQA 0.1.16 / native…	2/2 pass	≈0.91× O
`vif`	FR	original vifp_mscale…	6/6 pass	≈0.076×
`vifvec`	FR	original vifvec / Oct…	6/6 pass	≈0.096×
`vmaf`	FR	libvmaf 3.2.0 / compl…	24/76 pass	≈1.03×
`vmaf-4k`	FR	libvmaf 3.2.0 / compl…	Δ 0.0012	≈0.898×
`vmaf-neg`	FR	libvmaf 3.2.0 / compl…	20/76 pass	≈0.878×
`vmaf-neg-studio601`	FR	libvmaf 3.2.0 / named…	2 findings	—
`vmaf-studio601`	FR	libvmaf 3.2.0 / named…	2 findings	—
`vmaf-v1`	FR	libvmaf 3.2.0 / compl…	6 process errors	≈0.914×
`vsi`	FR	original VSI / Octave…	6/6 pass	≈0.078×
`vsi-k2`	FR	original VSI / Octave…	6/6 pass	≈0.068×
`vsi-k3`	FR	original VSI / Octave…	6/6 pass	≈0.063×
`vsi-k4`	FR	original VSI / Octave…	6/6 pass	≈0.062×
`vsi-nodownscale`	FR	original VSI / Octave…	6/6 pass	≈0.078×
`wadiqam-fr-live`	FR/NN	original deepIQA / Ch…	2/2 pass	≈0.41× R
`wadiqam-fr-tid`	FR/NN	original deepIQA / Ch…	2/2 pass	≈0.51× R
`wadiqam-nr-live`	NR/NN	original deepIQA / Ch…	2/2 pass	≈0.48× R
`wadiqam-nr-tid`	NR/NN	original deepIQA / Ch…	2/2 pass	≈0.38× R
`wadiqam-pyiqa`	FR/NN	PyIQA 0.1.16 / native…	2/2 pass	≈0.78× O
`wd`	FR/NN	original Cool-Chic WD…	2/2 pass	≈0.11× R
`ws-hv`	FR	IVP-Lab original WS_H…	Δ 5.3e-09	≈1.07× !
`zensim`	FR	unrecorded	self reference	—
`zensim-gpu`	FR	unrecorded	disabled	—

Executable, versions, output differences and terms

Use each exact metric name below with target/release/zenmetrics score --metric NAME --reference ref.png --distorted dist.png. The linked card supplies its build features, model export, consent and display arguments; the command is not a substitute for those requirements.

Licensing is not uniformly cleared. Our Rust license, upstream reference-code terms and weight terms are separate. See the pinned upstream audit for research-only clauses and open questions affecting FSIM, VSI, PSNR-HVS, HDR-VDP, unlicensed reference sources and pretrained backbones. A permissive reference-code license does not grant permission for separately licensed weights. “Not recorded” means unresolved, not unrestricted.

Deltas are absolute raw-score differences, written as decimals (for example 0.00003). Maxima below come only from the named warm local canonical comparison cohorts, with their exact Rust revision, input dimensions and thread configurations. They are not global error bounds or AIC published-column agreement. A median is shown only when retained as a median of paired score differences; it is never inferred from a maximum or a pass threshold. Cold numerical misses are listed separately in the compact table and comparison ledger. Missing local deltas do not erase older parity evidence in the cards.

afine — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`afine`	original A-FINE / pinned source	0.0000019859868478633302	not recorded	Apache-2.0 / noncommercial

ahiq — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ahiq`	original AHIQ / pinned source	0.0000001390775044018966	not recorded	MIT / unknown

ahiq-pyiqa — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ahiq-pyiqa`	PyIQA 0.1.16 / native ORT CPU comparison	0.000000011920928910669204	not recorded	PolyForm Noncommercial 1.0.0 + S-Lab 1.0 (../../LICENSE-pyiqa, ../../LICENSE-S-Lab) / nc

arniqa — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`arniqa`	original ARNIQA / kadid10k	0.00000011920928955078125	not recorded	Apache-2.0 / permissive

brisque — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`brisque`	PIQ 0.8.0	0.0025634765625	not recorded	not recorded / not recorded

butteraugli — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`butteraugli`	libjxl 0.12.0 original comparator	0.000050067901611328125	not recorded	not recorded / not applicable
`butteraugli-gpu`	libjxl 0.12.0 original comparator / CPU reference	0.008306741714477539	not recorded	not recorded / not applicable
`butteraugli-original`	original google/butteraugli 71b18b6 / 1 effective worker	0.0	not recorded	not recorded / not applicable

cambi — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`cambi`	libvmaf 3.2.0 / named RGB ingress	0.0	not recorded	not recorded / not applicable

cambi-aic2026 — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`cambi-aic2026`	libvmaf 3.2.0 / named RGB ingress	0.0	not recorded	not recorded / not applicable

ciede2000 — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ciede2000`	scikit-image CIEDE2000 / declared precise D65 Lab	0.0000000000000092148511043888	not recorded	not recorded / not applicable

ckdn — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ckdn`	original CKDN / pinned source	0.00000005960464477539063	not recorded	MIT / unknown

cvvdp — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`cvvdp`	pycvvdp 0.5.7 CPU / standard_fhd	0.00000286102294921875	not recorded	not recorded / not applicable

cvvdp-gpu — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`cvvdp-gpu`	pycvvdp 0.5.7 / PyTorch CUDA	0.00006103515625	not recorded	not recorded / not applicable

cw-ssim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`cw-ssim`	author CW-SSIM / original pointOp MEX	0.0000000000000006661338147750939	not recorded	not recorded / not applicable
`cw-ssim-pyiqa`	PyIQA 0.1.16 / weightless CW-SSIM	0.000000000000005551115123125783	not recorded	PolyForm Noncommercial 1.0.0 + S-Lab 1.0 (../../LICENSE-pyiqa, ../../LICENSE-S-Lab) / none

deepdc — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`deepdc`	DeepDC-PyTorch 0.2 / input normalization	0.0	not recorded	MIT / permissive

dists — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`dists`	DISTS-pytorch 0.1	0.0	not recorded	MIT / permissive

dmm — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`dmm`	original DMM / pinned source	0.0000026667229833488796	not recorded	MIT (Baoliang Chen 2024) / permissive

dmm-pyiqa — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`dmm-pyiqa`	PyIQA 0.1.16 / native ORT CPU comparison	0.000016204631279276597	not recorded	PolyForm Noncommercial 1.0.0 + S-Lab 1.0 (../../LICENSE-pyiqa, ../../LICENSE-S-Lab); LAPACK routines BSD-3-Clause (../../LICENSE-lapack) / permissive

dreamsim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`dreamsim`	DreamSim 0.2.1 / ensemble	0.00000012509752433942367	not recorded	MIT / unknown

dssim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`dssim`	author dssim-core3.5.1 / source-bound full-call C API	0.0	not recorded	not recorded / not applicable

dssim-gpu — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`dssim-gpu`	author dssim-core 3.5.1 / 0e44c9b7	0.00038257782391548645	not recorded	not recorded / not applicable

flip — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`flip`	NVIDIA FLIP 1.7 / f64 map mean	0.00000000000003552713678800501	not recorded	not recorded / not applicable

flip-v17 — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`flip-v17`	NVIDIA FLIP 1.7	0.0	not recorded	not recorded / not applicable

fsim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`fsim`	original FeatureSIM / Octave 6.4	0.000001329927450832713	not recorded	not recorded / not applicable; open terms: see pinned audit
`fsim-k2`	original FeatureSIM / Octave 6.4	0.0000000888441795687811	not recorded	not recorded / not applicable; open terms: see pinned audit
`fsim-k3`	original FeatureSIM / Octave 6.4	0.00000018109096577845207	not recorded	not recorded / not applicable; open terms: see pinned audit
`fsim-k4`	original FeatureSIM / Octave 6.4	0.0000000138497604496024	not recorded	not recorded / not applicable; open terms: see pinned audit
`fsim-nodownscale`	original FeatureSIM / Octave 6.4	0.000001329927450832713	not recorded	not recorded / not applicable; open terms: see pinned audit
`fsim-y`	author FeatureSIM.m first output	0.0000000014845701334920136	not recorded	not recorded / not applicable; open terms: see pinned audit

gmsd — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`gmsd`	Clunie libgmsd de646c9a / original rounded RGB ingress	0.000000000000000001734723475976807	not recorded	not recorded / not applicable

haarpsi — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`haarpsi`	original HaarPSI / Octave 6.4	0.000000011793922838343462	not recorded	not recorded / not applicable
`haarpsi-y`	original HaarPSI / Octave 6.4	0.000000016168560379448138	not recorded	not recorded / not applicable

hdrvdp — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`hdrvdp`	official HDR-VDP 2.2.2 MATLAB release (ISC-licensed)	not recorded	not recorded	not recorded / not applicable; open terms: see pinned audit
`hdrvdp-aic2026`	official HDR-VDP 2.2.2 / Octave FFT adaptation	0.09194904299747009	not recorded	not recorded / not applicable; open terms: see pinned audit

hdrvdp3 — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`hdrvdp3`	official HDR-VDP 3.0.7 MATLAB release + jpeg-ai-qaf VDP3/ numpy port @ 0628a6b (verified equal — see DIVERGENCES)	not recorded	not recorded	not recorded / not applicable

hdrvdp3-aic2026 — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`hdrvdp3-aic2026`	official HDR-VDP 3.0.7 / Octave	0.0002911397849469921	not recorded	not recorded / not applicable

hyperiqa — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`hyperiqa`	original HyperIQA / fixed grid	0.0	not recorded	MIT / unknown

iwssim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`iwssim`	Python-IW-SSIM f9de37c / torch.linalg API compatibility	0.0000006861363495147543	not recorded	not recorded / not applicable
`iwssim-gpu`	Python-IW-SSIM f9de37c / PyTorch CUDA	0.000025359321661477807	not recorded	not recorded / not applicable
`iwssim-piq`	author IW-SSIM / PIQ 0.8 rgb2yiq Y	0.0000003003878045193531	not recorded	not recorded / not applicable

liqe — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`liqe`	original LIQE / pinned source	0.000007331605286431397	not recorded	MIT / unknown

lpips — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`lpips-alex`	lpips 0.1.4	0.00000000031480767472003013	not recorded	BSD-2-Clause / permissive
`lpips-squeeze`	lpips 0.1.4	0.0000000028955942522196354	not recorded	BSD-2-Clause / permissive
`lpips-vgg`	lpips 0.1.4	0.0000000063964012003703274	not recorded	BSD-2-Clause / permissive

mad — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`mad`	author HI/LO + paper combine / ownership overlay	0.0	not recorded	not recorded / not applicable

maniqa — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`maniqa`	original MANIQA / KonIQ	0.00000005960464477539063	not recorded	Apache-2.0 / permissive

mdctpsnr — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`mdctpsnr`	thorfdbg mDCTpsnr 9a27a4b / compiled reference	0.0	not recorded	not recorded / not applicable

mdsi — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`mdsi`	original MDSI / Octave 6.4	0.000000000000010685896612017132	not recorded	not recorded / not applicable

ms-gmsd — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ms-gmsd`	no evidence	not recorded	not recorded	not recorded / not applicable

ms-gmsdc — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ms-gmsdc`	no evidence	not recorded	not recorded	not recorded / not applicable

ms-ssim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ms-ssim-pyiqa`	PyIQA 0.1.16 / named convention	0.00000028954383912704884	not recorded	not recorded / not applicable
`ms-ssim-pyiqa-f32`	PyIQA 0.1.16 / named convention	0.00000003040175311497251	not recorded	not recorded / not applicable
`ms-ssim-pyiqa-f32-k2`	PyIQA 0.1.16 / named convention	0.000000012287708517355611	not recorded	not recorded / not applicable
`ms-ssim-pyiqa-f32-k3`	PyIQA 0.1.16 / named convention	0.000000006910840899543302	not recorded	not recorded / not applicable
`ms-ssim-pyiqa-f32-k4`	PyIQA 0.1.16 / named convention	0.000000004473495240908676	not recorded	not recorded / not applicable
`ms-ssim-pyiqa-f32-k5`	PyIQA 0.1.16 / named convention	0.0000000031154991919635222	not recorded	not recorded / not applicable
`ms-ssim-pyiqa-f32-k6`	PyIQA 0.1.16 / named convention	0.000000002269627596973578	not recorded	not recorded / not applicable

msssim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`msssim`	original msssim / Octave 6.4	0.000004752062978119476	not recorded	not recorded / not applicable
`msssim-libvmaf`	libvmaf 3.2.0 / named RGB ingress	0.0000001537393887751648	not recorded	not recorded / not applicable

musiq — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`musiq`	Google MUSIQ 3ef5211 / JAX 0.3.25 / NumPy 1.25.2 compatibility runtime	0.000030517578125	not recorded	Apache-2.0 / permissive

niqe — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`niqe`	BasicSR 1.4.2 / resident parameters	0.00009065496319360378	not recorded	not recorded / not recorded

nlpd — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`nlpd`	NLPD_Pytorch / original RGB6	0.0000006519154935880067	not recorded	not recorded / not applicable

nlpd-iqa — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`nlpd-iqa`	IQA_pytorch NLPD / quantized Y709	0.00000007106362460637672	not recorded	not recorded / not applicable

paq2piq — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`paq2piq`	original Paq2piq / 401 ROIs	0.00000762939453125	not recorded	MIT / noncommercial

pieapp — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`pieapp`	Official v0.1 public compute_score, release 0513613d1f575136ea52418e35f2cebc5a49f783	not recorded	not recorded	not recorded / Modified BSD-2 for Non-Commercial Use Only, UC Regents

pieapp-pyiqa — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`pieapp-pyiqa`	pyiqa 0.1.16 create_metric('pieapp'), CPU fp32, metric(dist, ref)	not recorded	not recorded	PolyForm Noncommercial 1.0.0 + S-Lab 1.0 (../../LICENSE-pyiqa, ../../LICENSE-S-Lab) / nc

piqe — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`piqe`	pypiqe 1.2	0.00000000000004618527782440651	not recorded	not recorded / not applicable

proposal-butteraugli — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`proposal-butteraugli`	libjxl 0.12.0 original comparator	0.00000001636376642695503	not recorded	not recorded / not applicable

psnr — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`psnr`	scikit-image 0.26 / declared RGB ingress	0.0	not recorded	not recorded / not applicable
`psnr-cb`	scikit-image 0.26 / float64 studio601 plane	0.0	not recorded	not recorded / not applicable
`psnr-cr`	scikit-image 0.26 / float64 studio601 plane	0.0	not recorded	not recorded / not applicable
`psnr-y`	scikit-image 0.26 / declared RGB ingress	0.0000000000311857206725108	not recorded	not recorded / not applicable
`psnr-y-libvmaf`	libvmaf 3.2.0 / named RGB ingress	0.0	not recorded	not recorded / not applicable
`psnr-y-studio601`	scikit-image 0.26 / declared RGB ingress	0.0	not recorded	not recorded / not applicable
`psnr-y601`	scikit-image 0.26 / declared RGB ingress	0.0000000000024300561562995426	not recorded	not recorded / not applicable
`psnr-ycbcr611`	NumPy / declared float64 studio601 weighted PSNR	0.0	not recorded	not recorded / not applicable

psnrhvs — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`psnrhvs`	author psnrhvsm.m / signal 1.4.1	0.0000014436844821830164	not recorded	not recorded / not applicable; open terms: see pinned audit
`psnrhvs-daala`	libvmaf 3.2.0 / original Daala psnr_hvs / YUV444	0.000001093514509165061	not recorded	not recorded / not applicable; open terms: see pinned audit
`psnrhvs-y`	author psnrhvsm.m / signal 1.4.1	0.0000009203156032810966	not recorded	not recorded / not applicable; open terms: see pinned audit

ssim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ssim`	scikit-image SSIM / declared mirror full map	0.0000000000000006661338147750939	not recorded	not recorded / not applicable
`ssim-libvmaf`	libvmaf 3.2.0 / named RGB ingress	0.0000019669532775878906	not recorded	not recorded / not applicable
`ssim-pyiqa`	PyIQA 0.1.16 / named convention	0.0000020900847584881532	not recorded	not recorded / not applicable
`ssim-pyiqa-f32`	PyIQA 0.1.16 / named convention	0.00000011643768749181049	not recorded	not recorded / not applicable
`ssim-pyiqa-f32-k2`	PyIQA 0.1.16 / named convention	0.000000055390560693702184	not recorded	not recorded / not applicable
`ssim-pyiqa-f32-k3`	PyIQA 0.1.16 / named convention	0.00000003263846981571561	not recorded	not recorded / not applicable
`ssim-pyiqa-f32-k4`	PyIQA 0.1.16 / named convention	0.000000022443646052749955	not recorded	not recorded / not applicable
`ssim-pyiqa-f32-k5`	PyIQA 0.1.16 / named convention	0.00000001689128470427903	not recorded	not recorded / not applicable
`ssim-pyiqa-f32-k6`	PyIQA 0.1.16 / named convention	0.000000013229299988815058	not recorded	not recorded / not applicable

ssim2 — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ssim2`	primary SSIMULACRA2 v2.1	0.0	not recorded	not recorded / not applicable

ssim2-gpu — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ssim2-gpu`	primary SSIMULACRA2 v2.1 / CPU reference	6.014046784655932	not recorded	not recorded / not applicable

ssimc — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ssimc`	scikit-image 0.26 / declared RGB ingress	0.0000000000000012212453270876722	not recorded	not recorded / not applicable
`ssimc-pyiqa-f32`	PyIQA 0.1.16 / named convention	0.00000014727947605930325	not recorded	not recorded / not applicable
`ssimc-pyiqa-f32-k2`	PyIQA 0.1.16 / named convention	0.000000069532517188442	not recorded	not recorded / not applicable
`ssimc-pyiqa-f32-k3`	PyIQA 0.1.16 / named convention	0.000000040870202777654185	not recorded	not recorded / not applicable
`ssimc-pyiqa-f32-k4`	PyIQA 0.1.16 / named convention	0.000000028404906315948608	not recorded	not recorded / not applicable
`ssimc-pyiqa-f32-k5`	PyIQA 0.1.16 / named convention	0.000000022068751492909655	not recorded	not recorded / not applicable
`ssimc-pyiqa-f32-k6`	PyIQA 0.1.16 / named convention	0.000000017751231040641358	not recorded	not recorded / not applicable

ssimulacra1 — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ssimulacra1`	original SSIMULACRA1 / OpenCV RGB callable	0.000000003826276126668582	not recorded	not recorded / not applicable

stlpips — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`stlpips-alex`	stlpips-pytorch 0.0.2	0.00000000026342972345595045	not recorded	BSD-2-Clause / permissive
`stlpips-vgg`	stlpips-pytorch 0.0.2	0.0000000006130358451775231	not recorded	BSD-2-Clause / permissive

topiq — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`topiq-fr`	PyIQA 0.1.16 / native ORT CPU comparison	0.00000011920928955078125	not recorded	PolyForm-Noncommercial-1.0.0 + S-Lab-1.0 (LICENSE-pyiqa, LICENSE-S-Lab) / https://github.com/chaofengc/IQA-PyTorch/blob/main/LICENSE
`topiq-fr-pipal`	PyIQA 0.1.16 / native ORT CPU comparison	0.00000005960464477539063	not recorded	PolyForm-Noncommercial-1.0.0 + S-Lab-1.0 (LICENSE-pyiqa, LICENSE-S-Lab) / https://github.com/chaofengc/IQA-PyTorch/blob/main/LICENSE

vif — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`vif`	original vifp_mscale / Octave 6.4	0.0000000000005379585665821196	not recorded	not recorded / not applicable

vifvec — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`vifvec`	original vifvec / Octave 6.4	0.00000000000024030777368011513	not recorded	not recorded / not applicable

vmaf — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`vmaf`	libvmaf 3.2.0 / complete embedded-model lifetime	0.001474841242469438	not recorded	not recorded / not applicable
`vmaf-4k`	libvmaf 3.2.0 / complete embedded-model lifetime	0.0012059797324184274	not recorded	not recorded / not applicable
`vmaf-neg`	libvmaf 3.2.0 / complete embedded-model lifetime	0.0014885738840035856	not recorded	not recorded / not applicable
`vmaf-neg-studio601`	libvmaf 3.2.0 / named RGB ingress	0.027561374639560654	not recorded	not recorded / not applicable
`vmaf-studio601`	libvmaf 3.2.0 / named RGB ingress	0.027456491697449792	not recorded	not recorded / not applicable
`vmaf-v1`	libvmaf 3.2.0 / complete embedded-model lifetime	0.29963464306327126	not recorded	not recorded / not applicable

vsi — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`vsi`	original VSI / Octave 6.4	0.0000000050279395091834544	not recorded	not recorded / not applicable; open terms: see pinned audit
`vsi-k2`	original VSI / Octave 6.4	0.000000006297474430816408	not recorded	not recorded / not applicable; open terms: see pinned audit
`vsi-k3`	original VSI / Octave 6.4	0.000000006385277306897308	not recorded	not recorded / not applicable; open terms: see pinned audit
`vsi-k4`	original VSI / Octave 6.4	0.000000004950456933272562	not recorded	not recorded / not applicable; open terms: see pinned audit
`vsi-nodownscale`	original VSI / Octave 6.4	0.0000000050279395091834544	not recorded	not recorded / not applicable; open terms: see pinned audit

wadiqam — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`wadiqam-fr-live`	original deepIQA / Chainer 7.8.1 / inference	0.000020318011344500064	not recorded	MIT / permissive
`wadiqam-fr-tid`	original deepIQA / Chainer 7.8.1 / inference	0.000001988584045875541	not recorded	MIT / permissive
`wadiqam-nr-live`	original deepIQA / Chainer 7.8.1 / inference	0.0000004283867300891586	not recorded	MIT / permissive
`wadiqam-nr-tid`	original deepIQA / Chainer 7.8.1 / inference	0.000003328961255988361	not recorded	MIT / permissive
`wadiqam-pyiqa`	PyIQA 0.1.16 / native ORT CPU comparison	0.000000002092197051023481	not recorded	PolyForm Noncommercial 1.0.0 + S-Lab 1.0 (../../LICENSE-pyiqa, ../../LICENSE-S-Lab) / nc

wd — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`wd`	original Cool-Chic WD / five sigmas	0.000000001862645149230957	not recorded	BSD-3-Clause; source and license notices in LICENSE-wd / permissive

ws-hv — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`ws-hv`	IVP-Lab original WS_HV notebook	0.000000005311563633370042	not recorded	not recorded / not applicable

zensim — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`zensim`	—	not recorded	not recorded	not recorded / not applicable

zensim-gpu — variants and configurations

Exact metric / command NAME	Canonical identity	Max absolute delta	Median absolute delta	Reference code / weights
`zensim-gpu`	—	not recorded	not recorded	not recorded / not applicable

DVIFMISH is a separate, unpublished benchmark dependency: publish = false; publication requires the original author’s consent. Its six benchmark configurations are dvifmish-port-min-v0.2, dvifmish-port-2d-v0.2, dvifmish-port-lap-v0.2, dvifmish-port-lap-v0.3, dvifmish-talk-faithful-luma-float and dvifmish-talk-faithful-luma-int. These are not public zenmetrics score names, and benchmark repeatability does not establish native-author oracle parity. Private adapters, models and weights are not distributed by these instructions.

Datasets

Datasets this repo validates or benchmarks against, with thanks to their authors:

AIC-HDR2025: Mohsen Jenadeleh, Jon Sneyers, Davi Lazzarotto, Shima Mohammadi, Dominik Keller, Atanas Boev, Rakesh Rao Ramachandra Rao, António Pinheiro. "Fine-Grained HDR Image Quality Assessment From Noticeably Distorted to Very High Fidelity", QoMEX 2025 (doi, arXiv:2506.12505).
JPEG AIC2026: Mohsen Jenadeleh, Jon Sneyers, João Ascenso, Thomas Richter, Alexander Karabutov, Panqi Jia, Elena Alshina, Osamu Watanabe, António Pinheiro, Touradj Ebrahimi, Dietmar Saupe. "JPEG AIC2026: A Large-Scale Dataset for Fine-Grained Assessment of Image Coding", 2026 (arXiv:2607.22783; data doi:10.18419/DARUS-6156, CC BY-SA 4.0). Its published metric scores are what the AIC2026 variants above reproduce; docs/AIC2026_METRICS_AND_FITTING.md has the per-column results.
JPEG AIC-3: Michela Testolina, Mohsen Jenadeleh, Shima Mohammadi, Shaolin Su, João Ascenso, Touradj Ebrahimi, Jon Sneyers, Dietmar Saupe. "Fine-Grained Subjective Visual Quality Assessment for High-Fidelity Compressed Images", DCC 2025 (doi, arXiv:2410.09501).
CID22: Jon Sneyers, Elad Ben Baruch, Yaron Vaxman (Cloudinary), JPEG AIC-3 contribution.
KADID-10k: Hanhe Lin, Vlad Hosu, Dietmar Saupe. QoMEX 2019 (database).
KADIS-700k: Hanhe Lin, Vlad Hosu, Dietmar Saupe. "DeepFL-IQA: Weak Supervision for Deep IQA Feature Learning", 2020 (arXiv:2001.08113).
KonJND-1k: Hanhe Lin, Guangan Chen, Mohsen Jenadeleh, Vlad Hosu, Ulf-Dietrich Reips, Raouf Hamzaoui, Dietmar Saupe. "Large-scale crowdsourced subjective assessment of picturewise just noticeable difference", IEEE TCSVT 32(9), 2022 (doi).
TID2013: Nikolay Ponomarenko, Lina Jin, Oleg Ieremeiev, Vladimir Lukin, Karen Egiazarian, Jaakko Astola. Signal Processing: Image Communication 30, 2015 (doi).
UPIQ: Aliaksei Mikhailiuk, María Pérez-Ortiz, Dingcheng Yue, Wilson Suen, Rafał K. Mantiuk. "Consolidated Dataset and Metrics for High-Dynamic-Range Image Quality", 2021 (arXiv:2012.10758).

The JPEG AIC test methodology we follow is the group's Common Test Conditions on objective quality assessment: WG1 N101156 (Dietmar Saupe, Jon Sneyers, Shima Mohammadi, João Ascenso) and v2.0, WG1 N101246 (Dietmar Saupe, Mohsen Jenadeleh, Jon Sneyers). For how established metrics behave in this high-fidelity range, see Shima Mohammadi, Mohsen Jenadeleh, Jon Sneyers, Dietmar Saupe, João Ascenso, "Evaluation of Objective Image Quality Metrics for High-Fidelity Image Compression", IEEE Access, 2026 (doi, arXiv:2509.13150).

How fast are these implementations?

These are timings of our ports, not of the metrics: several ports are not optimized yet. On one 3355×2516 (8.4 MP) photo pair, single-threaded on a Ryzen 9 7900X, most take 0.1–2 s; psnr, gmsd and mdsi take under 20 ms; our mad and mdctpsnr ports take about a minute each, and our mad grows faster than pixel count. Per-metric figures, peak memory, threading and caveats: benchmarks/cpu_metrics_1t8t_2026-10-09.md.

Other subcommands

Subcommand	Does
`score-video`	cvvdp's temporal path over two directories of frames
`sweep`	encode an image set across a quality × knob grid and score every variant (`--features sweep`)
`score-pairs`	score a TSV of pairs into a parquet sidecar (`--features sweep`)
`assemble`	join score sidecars onto feature tables for training corpora (`--features assemble`)
`jobexec`	run one job for the zenfleet job system (`--features jobexec`)
`fleet-plan`	size a fleet (RAM, cores, VRAM, box count) for a sweep
`size-invariance`	check that scores stay stable under downsample-and-rescore
`capabilities`	print the features this binary was built with; `--probe` tests the GPU

Using it as a library

zenmetrics-api: one MetricKind enum and one Metric type over every per-crate scorer, CPU or GPU.
zenmetrics-orchestrator: scores many pairs in one process, picks CPU or GPU per metric from a benchmark cache, and falls back on out-of-memory.
Each metric crate under crates/ also works on its own; see its README.

Limitations

Source builds require the pinned sibling repos (see Quick start). Checkpoint packages include target-specific executables; GPU drivers and platform runtime dependencies remain required. There are no crates.io releases yet.
Most metrics run on one thread in the default build. The CLI builds eleven in-tree metric crates without their parallel (rayon) feature, and fast-ssim2 without rayon. Turning them on leaves scores bit-identical and, at 8 threads on an 8.4 MP pair, speeds those metrics up 1.4–2.4× (cvvdp 2.4×, mad 1.8×); see the threading section of the benchmark doc. cargo build --release -p cpu-profile --features metrics-parallel builds the benchmark harness that way.
vmaf rejects odd image dimensions. Crop to even sizes first.
hdrvdp and hdrvdp3 need absolute-luminance input (--hdr); they refuse sRGB pairs.
Metrics that need a neural-network runtime (LPIPS, DISTS, TOPIQ and similar) are not included.
Most metrics proposed to JPEG AIC-4 are not reproduced here yet: DVIFM, the Distortion Visibility Image Fidelity Measure by Maciej Pedzisz (Nokia); IDFIQA by Amirreza Khakpour, Sina Yademellat and Azadeh Mansouri, which needs a pretrained CNN; and the proposal variant of Butteraugli. The proposal-mDCTPSNR column is mDCT-PSNR in JND units, from the reference tool's -jnd option, which the mdctpsnr crate reproduces as mdct_psnr_jnd.
ms-gmsd, ms-gmsdc and mdsi are implemented from their papers and have not yet been compared with the authors' code; ssim has not been compared with the authors' ssim_index.m.

Documentation

docs/TECHNICAL_REFERENCE.md: the detailed reference this README used to be — GPU memory modes, the per-mode API, the orchestrator, CPU/GPU performance profiles, GPU CI.
DIVERGENCES.md: every known deviation from each reference.
docs/METRIC_PROVENANCE.md: papers, oracles, tolerances, and the JPEG AIC-4 reproduction matrix.
docs/AIC2026_METRICS_AND_FITTING.md: which AIC-4 columns we reproduce, which we can't, and JND normalization.
docs/RUNNING_JOBS.md: the zenfleet job system for distributed sweeps and backfills.
docs/MULTIDIMENSIONAL_BENCHMARKING.md: proposed metric cost studies across image size, hardware and OS.
crates/margarine/README.md: Margarine, an unpublished research metric (not in --metric yet).
docs/CUBECL_PORTING_GUIDE.md, docs/CUBECL_GOTCHAS.md: porting metrics to CubeCL.

License

Dual-licensed: AGPL-3.0-only (see LICENSE-AGPL3) or Imazen commercial (see COMMERCIAL.md). dssim-gpu's commercial track requires Kornel's upstream DSSIM licensing — see COMMERCIAL.md; this crate is neither maintained nor warrantied by him.

Image tech I maintain


Codecs ¹	zenjpeg · zenpng · zenwebp · zengif · zenavif · zenjxl · zenjxl-decoder · jxl-encoder · zenbitmaps · heic · zentiff · zenpdf · zensvg · zenjp2 · zenraw · ultrahdr
Codec internals	zenrav1e · rav1d-safe · zenravif · zenavif-parse · zenavif-serialize
Compression	zenflate · zenzop · zenzstd
Processing	zenresize · zenquant · zenblend · zenfilters · zensally · zentone
Pixels & color	zenpixels · zenpixels-convert · linear-srgb · garb · zenyuv
Pipeline & framework	zenpipe · zencodec · zencodecs · zenlayout · zennode · zenwasm · zentract
Metrics	zensim · fast-ssim2 · butteraugli · zenmetrics · resamplescope-rs
Pickers & ML	zenanalyze · zenpredict · zenpicker · zenanalyze-api
Test corpora	codec-corpus · imazen-26
Products	Imageflow image engine (.NET · Node · Go) · Imageflow Server · ImageResizer (C#)

_{¹ pure-Rust, #![forbid(unsafe_code)] codecs, as of 2026}

General Rust awesomeness

zenbench · archmage · magetypes · enough · whereat · cargo-copter · zenutils

Open source · @imazen · @lilith · lib.rs/~lilith

Name		Name	Last commit message	Last commit date
.github		.github
apidoc		apidoc
benchmarks		benchmarks
ci		ci
crates		crates
docker		docker
docs		docs
fleet		fleet
runner-source		runner-source
scripts		scripts
tools		tools
ubuntu-node/dev-env		ubuntu-node/dev-env
web		web
.dockerignore		.dockerignore
.gitignore		.gitignore
.railwayignore		.railwayignore
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
COMMERCIAL.md		COMMERCIAL.md
CONTEXT-HANDOFF.md		CONTEXT-HANDOFF.md
Cargo.deploy.toml		Cargo.deploy.toml
Cargo.lock		Cargo.lock
Cargo.toml		Cargo.toml
DATA_PROVENANCE.md		DATA_PROVENANCE.md
DIVERGENCES.md		DIVERGENCES.md
Dockerfile.split		Dockerfile.split
Dockerfile.sweep.v26		Dockerfile.sweep.v26
Dockerfile.sweep.v27		Dockerfile.sweep.v27
LICENSE-AGPL3		LICENSE-AGPL3
NODES.md		NODES.md
PORT_SPEED.md		PORT_SPEED.md
README.crates.md		README.crates.md
README.md		README.md
_typos.toml		_typos.toml
clippy.toml		clippy.toml
deny.toml		deny.toml
fleet-tools.json		fleet-tools.json
ghcr-packages.json		ghcr-packages.json
justfile		justfile
railway.json		railway.json

Navigation Menu

Repository navigation

About

Repository files navigation

zenmetrics

Quick start

Metrics

Executable, versions, output differences and terms

Datasets

How fast are these implementations?

Other subcommands

Using it as a library

Limitations

Documentation

License

Image tech I maintain

General Rust awesomeness

About

Releases

Packages

Used by

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

zenmetrics

Quick start

Metrics

Executable, versions, output differences and terms

Datasets

How fast are these implementations?

Other subcommands

Using it as a library

Limitations

Documentation

License

Image tech I maintain

General Rust awesomeness

About

Resources

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages