Skip to content
/ codec-corpus Public

Corpus of test images for testing and validating codecs.

License

View license
0 stars 0 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

imazen/codec-corpus

BranchesTags

Repository files navigation

Codec Corpus

A curated collection of reference images for codec quality calibration, compression benchmarking, and format conformance testing. Maintained by Imazen.

Total repo size: ~600 MB committed, plus ~1 GB available via download scripts.

Quick Start

# Clone everything (~600 MB)
git clone https://github.com/imazen/codec-corpus.git

# Or clone just one dataset using sparse checkout
git clone --depth 1 --filter=blob:none --sparse \
  https://github.com/imazen/codec-corpus.git
cd codec-corpus
git sparse-checkout set clic2025

# Add more datasets later
git sparse-checkout add CID22 gb82-sc qoi-benchmark

Datasets at a Glance

Quality Calibration & Compression Research

Dataset Images Size Resolution License Best For
CLIC 2025 62 219 MB ~2048px long edge Unsplash High-res lossy quality calibration
CID22 250 94 MB 512×512 CC BY-SA 4.0 Perceptual quality training, diverse content
KADID-10k 81 25 MB 512×384 Pixabay IQA research reference images
GB82 25 9.6 MB 576×576 CC0 Compact photographic benchmarking
GB82-SC 10 2.9 MB Various (640–2940px) CC0 Screen content & screenshot compression
QOI Benchmark 15+ 39 MB+ Various (1313×2874–8008) CC0/PD/Mixed Web screenshots, icons, textures
Kodak (Legacy) 24 15 MB 768×512 Unrestricted Deprecated — historical reference only

Format Conformance & Edge Cases

Dataset Files Size License Best For
JPEG Conformance 177 6.7 MB MIT/IJG+BSD/Various JPEG decoder conformance & robustness
JXL 180 88 MB BSD-3-Clause JPEG XL decoder conformance
PNGSuite 176 720 KB Freeware PNG decoder conformance
image-rs 126 4.5 MB MIT Multi-format edge cases
zune-image 3,429 33 MB MIT/Apache-2.0/Zlib Fuzz testing, decoder robustness
mozjpeg 15 1.2 MB IJG + BSD JPEG codec reference files
imageflow 45 7.8 MB Various Orientation, format conversion edge cases

Quality Calibration & Compression Research

CLIC 2025

Challenge on Learned Image Compression 2025 — High-resolution photographic images curated for compression quality research. This is the most relevant modern benchmark for lossy codec calibration, featuring large, diverse photographs at resolutions representative of modern camera output and web delivery.

Folder Images Size Purpose
clic2025/training/ 32 103 MB Tune and calibrate encoders against these
clic2025/final-test/ 30 116 MB Holdout set — final evaluation only

Note: The CLIC challenge calls these "validation" and "test" respectively. We renamed "validation" → "training" because in codec development, "validation" typically implies a holdout set, which is the opposite of the intended use. See clic2025/README.md for the original naming and download links.

  • Resolution: ~2048px on long edge (varies from 878px to 2048px on short edge)
  • Format: Lossless PNG, 8-bit sRGB
  • Source: https://clic2025.compression.cc/
  • License: Unsplash License — Free for any use, no attribution required, cannot be sold unmodified or used to build a competing service.

CID22

Cloudinary Image Dataset 2022 — 250 diverse images selected by Cloudinary for perceptual quality research. This dataset spans a wider variety of content types than most photographic benchmarks: portraits, landscapes, text, graphics, medical imagery, scientific plots, and more. The 512px size makes it fast to process while remaining large enough for meaningful perceptual quality evaluation.

For compression benchmarking, CID22 is one of the best choices available: it was specifically designed for this purpose, its diversity avoids the over-fitting that plagues small homogeneous corpora, and the training/validation split enables principled evaluation.

Folder Images Purpose
CID22/CID22-512/validation/ 41 Held out for validation
CID22/CID22-512/training/ 209 Model training and calibration

KADID-10k

Konstanz Artificially Distorted Image quality Database — 81 pristine reference images used for image quality assessment (IQA) research. Created at the University of Konstanz, this dataset is widely used for training and evaluating perceptual quality metrics. Only the pristine reference images are included here; the 10,125 distorted variants are not.

Folder Images Size
kadid10k/ 81 25 MB
  • Resolution: 512×384, 8-bit sRGB
  • Format: Lossless PNG
  • Source: https://database.mmsp-kn.de/kadid-10k-database.html
  • License: Pixabay License — Free for commercial and non-commercial use, no attribution required.
  • Citation: H. Lin, V. Hosu and D. Saupe, "KADID-10k: A Large-scale Artificially Distorted IQA Database," 2019 Eleventh International Conference on Quality of Multimedia Experience (QoMEX), Berlin, Germany, 2019, pp. 1-3, doi: 10.1109/QoMEX.2019.8743252.

GB82

GB82 Image Dataset by Gianni Rosato — A compact, purpose-built CC0 dataset for image compression algorithm development. The 25 images are selected to be challenging: fine facial details, low-contrast sky gradients, digital noise, fine textures. Optimizing for weak metrics like PSNR should not yield visually compelling results on most images.

Folder Images Size
gb82/ 25 9.6 MB

Image categories:

  • 3 portraits (2 human, 1 dog)

  • 6 landscapes

  • 8 closeups of inanimate objects or scenes

  • 4 low-light shots

  • 3 realistically rendered graphics

  • 1 non-photographic image

  • Resolution: 576×576, 8-bit sRGB, D65 white point

  • Source: https://github.com/gianni-rosato/gb82-image-set

  • License: CC0 1.0 — Public domain, no restrictions.

GB82-SC

GB82 Screen Content Dataset by Gianni Rosato — Screenshots and screen content images focusing on text, UI elements, and graphics from five platforms. Screen content compresses very differently from photographic content: sharp edges, flat color regions, anti-aliased text, and high-contrast UI elements stress different codec paths. This dataset fills an important gap that photographic benchmarks miss entirely.

File Resolution Content
codec_wiki.png 2560×1664 Wikipedia article with text and diagrams
gmessages.png 1440×3088 Google Messages (Android)
graph.png 796×481 Data visualization / chart
gui.png 1356×1132 Desktop GUI elements
imac_dark.png 2940×1912 macOS desktop (dark mode)
imac_g3.png 2940×1912 macOS desktop (light, iMac G3 style)
imessage.png 1206×2622 iMessage conversation (iOS)
terminal.png 1646×1062 Terminal / command line
windows95.png 640×480 Windows 95 desktop
windows.png 2560×1392 Modern Windows desktop

QOI Benchmark Suite

Subsets from the QOI Benchmark Suite by Dominic Szablewski. The screenshot_web subset is committed directly to the repo; all other subsets can be fetched with the included download scripts.

Committed:

Subset Files Size License Content
qoi-benchmark/screenshot_web/ 14 39 MB CC0 1.0 Full-page web screenshots (amazon, apple, cnn, wikipedia, reddit, etc.)

Available via download scripts:

Subset Files Size License Content
icon_512 214 12 MB Public Domain Tango Icon Library at 512px
icon_64 214 1.3 MB Public Domain Tango Icon Library at 64px
screenshot_game 619 256 MB CC BY-SA 3.0 Game screenshots from Wikimedia Commons
textures_pk 1004 44 MB — Texture pack
textures_pk01 115 19 MB — Texture pack 01
textures_pk02 237 99 MB — Texture pack 02
textures_plants 61 50 MB — Plant textures
textures_photo 21 37 MB — Photographic textures
photo_kodak 25 15 MB Unrestricted Kodak suite (duplicate)
photo_tecnick 101 228 MB — Tecnick photographic set
photo_wikipedia 50 85 MB — Wikipedia photographs
pngimg 189 220 MB CC BY-NC 4.0 PNG images (non-commercial only) 
# Download all subsets (~1.1 GB tarball)
cd qoi-benchmark && ./download.sh

# Download specific subsets
./download.sh icon_512 icon_64

# List available subsets
./download.sh --list
# Windows
cd qoi-benchmark
.\download.ps1
.\download.ps1 -Subsets icon_512,icon_64
.\download.ps1 -List
  • Source: https://qoiformat.org/benchmark/
  • License: Per-subset (see table above). screenshot_web is CC0. icon_* subsets are Public Domain (Tango Icon Library). screenshot_game is CC BY-SA 3.0 (Wikimedia Commons). pngimg is CC BY-NC 4.0 (non-commercial only). Other subsets have unspecified licensing in the archive.

Kodak (Legacy)

⚠️ Do not use Kodak for new projects. This dataset is retained only for historical comparison with older research. See recommendations below.

Kodak Lossless True Color Image Suite — 24 images that were the de facto standard benchmark for image compression research from the 1990s through the 2010s.

Folder Images Size
kodak-legacy/ 24 15 MB
  • Resolution: 768×512 (or 512×768), 8-bit sRGB
  • Format: Lossless PNG
  • Source: http://r0k.us/graphics/kodak/
  • License: Unrestricted usage (released by Eastman Kodak Company)

Why Kodak is deprecated:

  • Overfit benchmark: Decades of codec tuning against these 24 images have made Kodak scores nearly meaningless. Many codecs are specifically optimized for Kodak, so performance on Kodak does not predict real-world quality.
  • Resolution mismatch: 768×512 is far smaller than modern display resolutions, camera output, or typical web images. Compression behavior at this size is not representative of actual usage.
  • Content bias: 24 images of mostly pastoral outdoor scenes circa 1990 do not reflect modern content: smartphone photos, social media, text overlays, UI screenshots, HDR, etc.
  • No holdout set: With only 24 images and no train/test split, any tuning against Kodak is also evaluation against Kodak.

Use instead:

Use Case Recommended Dataset
Fixed-size comparison (512×512) CID22 — 250 diverse images with training/validation split
Variable-size / high-res photos CLIC 2025 — modern high-res photographs (~2048px)
Compact photographic benchmarks GB82 — 25 challenging CC0 images at 576×576
Screenshots & screen content GB82-SC and QOI screenshot_web

Screenshot testing is essential for any codec deployed on the web. Text, UI elements, and graphics compress very differently from photographs — ignoring screen content leads to poor real-world performance.

Format Conformance & Edge Cases

JPEG Conformance

JPEG Decoder Conformance Test Suite — Files organized by expected decoder behavior, designed for systematic testing of JPEG decoders.

Folder Files Expected Behavior
jpeg-conformance/valid/ 41 MUST decode correctly
jpeg-conformance/invalid/ 116 MUST reject gracefully
jpeg-conformance/non-conformant/ 20 MAY reject or recover

valid/ — Reference JPEG images including camera samples from 12 manufacturers, restart intervals, CMYK/YCCK color spaces, and various sampling configurations.

invalid/ — Crash tests and malformed files from imagetestsuite and fuzzing. Decoders must not crash or hang on these.

non-conformant/ — Files that violate the JPEG spec but are common in the wild:

  • truncated/ — Files cut at various stream positions
  • extraneous-data/ — Extra bytes in unexpected locations
  • marker-quirks/ — Unusual marker sequences (e.g., multiple 0xFF before EOI)
  • metadata-quirks/ — ICC profile chunk issues (ordering, duplicates, missing chunks)
  • progressive-quirks/ — Progressive scan encoding edge cases

Each non-conformant file has a companion .txt file explaining the defect and expected strict vs. lenient decoder behavior.

JXL

JPEG XL Test Suite — Comprehensive coverage of JPEG XL decoder features and conformance.

Folder Files Size Purpose
jxl/conformance/ 39 6.2 MB Official libjxl conformance tests
jxl/features/ 128 81 MB Feature-specific test images
jxl/edge-cases/ 13 332 KB Minimal and boundary-condition tests

Feature coverage:

  • Encoding: Lossless (modular), lossy (VarDCT), progressive

  • Color: 8/12/16/32-bit depth, sRGB, linear, grayscale, CMYK

  • HDR: PQ and HLG transfer functions

  • Animation: Multi-frame, variable timing, splines

  • Alpha: Premultiplied, non-premultiplied, blend modes

  • Features: EXIF orientation, patches, ICC profiles, JPEG reconstruction

  • Source: https://github.com/libjxl/libjxl, https://github.com/libjxl/conformance

  • License: BSD-3-Clause

PNGSuite

Official PNG Conformance Test Suite by Willem van Schaik — Covers all PNG features for decoder conformance testing.

Folder Files Size
pngsuite/ 176 720 KB

Coverage:

  • Bit depths: 1, 2, 4, 8, 16

  • Color types: grayscale, RGB, palette, grayscale+alpha, RGBA

  • Interlacing (Adam7), transparency (tRNS), gamma correction

  • Ancillary chunks: sRGB, iCCP, tEXt, sBIT, pHYs, etc.

  • Corrupted files for error handling

  • Source: http://www.schaik.com/pngsuite/

  • License: Freeware — free to use, copy, modify, and distribute.

image-rs

Rust image library test images — Multi-format edge cases and malformed files used by the image-rs crate.

Folder Files Content
image-rs/test-images/bmp/ 60 BMP format variants and malformed files
image-rs/test-images/gif/ 11 GIF animation edge cases
image-rs/test-images/ico/ 7 Icon format tests
image-rs/test-images/jpg/ 7 JPEG metadata, progressive encoding
image-rs/test-images/png/ 22 16-bit, APNG, transparency
image-rs/test-images/tiff/ 10 TIFF compression, predictors
image-rs/test-images/webp/ 9 WebP lossless/lossy variants

zune-image

zune-image test suite — Fuzz corpus and decoder robustness tests from the zune-image project.

Folder Files Purpose
zune/test-images/jpeg/ 30 JPEG edge cases (CMYK, progressive, subsampling)
zune/fuzz-corpus/jpeg/ 1,836 Minimal JPEG fuzz inputs
zune/fuzz-corpus/png/ 837 Minimal PNG fuzz inputs
zune/fuzz-corpus/inflate/ 726 DEFLATE/inflate edge cases

The fuzz corpus files are minimal inputs designed to exercise specific code paths and edge cases in decoders. They are not meaningful images — they exist to catch crashes, hangs, and memory safety issues.

mozjpeg

Mozilla JPEG encoder test images — Reference files for JPEG codec testing, from the mozjpeg project.

File Purpose
testorig.ppm Source image (PPM format)
testorig.jpg Baseline JPEG reference
testimgari.jpg Arithmetic-coded JPEG
testimgint.jpg Progressive JPEG
testorig12.jpg 12-bit JPEG
shira_bird8.bmp, monkey16.ppm Additional source images
test.scan, test1.scan Custom scan scripts
test1.icc, test3.icc ICC color profiles

imageflow

Imageflow test inputs — Images used by imageflow for testing format conversion, orientation handling, and edge cases.

Folder Files Content
imageflow/test_inputs/ 29 WebP, JPEG, PNG, GIF test images
imageflow/test_inputs/orientation/ 16 EXIF orientation test set (all 8 orientations × landscape/portrait)

Includes: corrupt JPEG, color profile edge cases, transparency, gradients, whitespace handling, high-resolution (5760×4320) test image.

Directory Structure

codec-corpus/
├── clic2025/                    # CLIC 2025 (Unsplash License)
│   ├── LICENSE
│   ├── README.md                # Original naming & download links
│   ├── training/                # 32 high-res images (CLIC calls this "validation")
│   └── final-test/              # 30 high-res images (CLIC calls this "test")
├── CID22/                       # Cloudinary CID22 (CC BY-SA 4.0)
│   ├── LICENSE
│   └── CID22-512/
│       ├── validation/          # 41 images
│       └── training/            # 209 images
├── gb82/                        # GB82 photographic (CC0)
│   ├── LICENSE
│   └── *.png                    # 25 images
├── gb82-sc/                     # GB82 screen content (CC0)
│   ├── LICENSE
│   └── *.png                    # 10 screenshots
├── qoi-benchmark/               # QOI Benchmark Suite (CC0/PD/Mixed)
│   ├── README.md
│   ├── download.sh              # Fetch additional subsets (bash)
│   ├── download.ps1             # Fetch additional subsets (PowerShell)
│   └── screenshot_web/          # 14 web page screenshots (CC0)
├── kadid10k/                    # KADID-10k reference images (Pixabay License)
│   ├── LICENSE
│   └── *.png                    # 81 pristine images for IQA research
├── kodak-legacy/                # Kodak suite — DEPRECATED (Unrestricted)
│   ├── LICENSE
│   └── *.png                    # 24 classic images (historical reference only)
├── jxl/                         # JPEG XL tests (BSD-3-Clause)
│   ├── LICENSE
│   ├── conformance/             # 39 conformance tests
│   ├── features/                # 128 feature tests
│   └── edge-cases/              # 13 boundary tests
├── jpeg-conformance/            # JPEG conformance (MIT/IJG+BSD/Various)
│   ├── README.md
│   ├── SOURCES.md               # Per-file attribution
│   ├── valid/                   # 41 files — MUST decode
│   ├── invalid/                 # 116 files — MUST reject
│   └── non-conformant/          # 20 files — MAY reject (with .txt descriptions)
├── pngsuite/                    # PNG conformance (Freeware)
│   ├── LICENSE
│   └── *.png                    # 176 images
├── image-rs/                    # image-rs tests (MIT)
│   ├── LICENSE-MIT
│   └── test-images/
│       ├── bmp/, gif/, ico/
│       ├── jpg/, png/
│       ├── tiff/, webp/
├── zune/                        # zune-image tests (MIT/Apache-2.0/Zlib)
│   ├── LICENSE-MIT, LICENSE-APACHE, LICENSE-ZLIB
│   ├── test-images/jpeg/
│   └── fuzz-corpus/
│       ├── jpeg/, png/, inflate/
├── mozjpeg/                     # mozjpeg tests (IJG + BSD)
│   ├── LICENSE
│   └── *.ppm, *.jpg, *.bmp, *.icc
└── imageflow/                   # imageflow tests (Various)
    └── test_inputs/
        ├── orientation/         # 16 EXIF orientation tests
        └── *.jpg, *.png, *.webp, *.gif

License Summary

Every dataset includes its own license file in its directory.

Dataset License Commercial Use Attribution Required ShareAlike
CLIC 2025 Unsplash License Yes No No
CID22 CC BY-SA 4.0 Yes Yes Yes
GB82 CC0 1.0 Yes No No
GB82-SC CC0 1.0 Yes No No
QOI screenshot_web CC0 1.0 Yes No No
QOI icon_* Public Domain Yes No No
QOI screenshot_game CC BY-SA 3.0 Yes Yes Yes
QOI pngimg CC BY-NC 4.0 No Yes No
KADID-10k Pixabay License Yes No No
Kodak (Legacy) Unrestricted Yes No No
JXL BSD-3-Clause Yes No No
PNGSuite Freeware Yes No No
image-rs MIT Yes No No
zune-image MIT/Apache-2.0/Zlib Yes No No
mozjpeg IJG + BSD Yes No No
JPEG Conformance MIT/IJG+BSD/Various Yes Varies No
imageflow Various Yes Varies No

Choosing a Dataset

For lossy codec quality calibration: Use CLIC 2025 (high-res, modern photos) and CID22 (diverse content, training/validation split). These are the most representative and methodologically sound choices for modern codec evaluation.

For compact benchmarking: Use GB82 (25 challenging photos, CC0, 576×576). Fast to process, explicitly designed to resist metric gaming.

For screen content / non-photographic images: Use GB82-SC and QOI Benchmark screenshot_web. Screenshots, UI elements, text, and graphics compress very differently from photos — testing both content types is essential for any codec deployed on the web.

For decoder conformance: Use the format-specific test suites: JPEG Conformance, JXL, PNGSuite. For fuzz/robustness testing, use zune-image.

Do not use Kodak for modern codec evaluation. It is retained only for historical comparison — see the deprecation note. For any new benchmarking work, use CID22 (fixed 512×512) or CLIC 2025 (variable high-res), and always include screenshot/screen content testing via GB82-SC or QOI screenshot_web.

Contributing

To suggest additional datasets, please open an issue with:

  • Source URL
  • License information
  • Description of what the dataset tests or what content type it represents

About

Corpus of test images for testing and validating codecs.

Resources

Readme

License

View license
Activity
Custom properties

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages