Hugging Face's logo

Datasets:
yifeipet
/
AFP-GIC-Reconstructions-with-Metrics

Size:
1K<n<10K
ArXiv:
Tags:
image-compression
generative-image-compression
controllable-image-compression
image-reconstruction
kodak
clic2020
Dataset card Files
xet
 Community
The Dataset Viewer has been disabled on this dataset.

AFP-GIC Reconstructed Images and Evaluation Metrics

2,760 reconstructed images. Three benchmarks. Five bitrate operating points.

This is the official reconstruction and evaluation-data release for AFP-GIC, a content-adaptive generative image compression model supporting five controllable bitrate operating points within one pretrained model. It accompanies Adaptive Fused Prior Transfer for Controllable Generative Image Compression, by Yifei Pei, Ying Liu, and Nam Ling, published in IEEE Access (2026).

Download reconstructed RGB PNG images, per-image metrics, and dataset-level summaries to include AFP-GIC in research comparisons under matched evaluation protocols, without rerunning the pretrained model.

Coverage

Dataset Evaluated split Source images Operating points Reconstructions
Kodak Complete 24-image benchmark 24 5 120
CLIC2020 Test images used in the paper 428 5 2,140
DIV2K Validation images 100 5 500
Total 552 2,760

The five operating points use the same AFP-GIC pretrained model. The package contains the final model's reconstructions and evaluation records, rather than baseline reconstructions or additional ablation experiments. Original reference images, training images, and compressed bitstreams are not included.

Download and extract

Open Files and versions to download the archives directly. This repository distributes ZIP archives; it does not require a GPU or the AFP-GIC model to view the saved images and CSV files.

There are 11 independent image ZIP files. Download only the datasets and operating points you need. Also download AFP-GIC_Metrics_and_Documentation.zip for the consolidated CSV files and this documentation.

Image archive Contents
AFP-GIC_Kodak_Reconstructions_q0-q4.zip All 120 Kodak reconstructions
AFP-GIC_CLIC2020_Reconstructions_q0.zip through q4.zip Five independent archives, 428 images each
AFP-GIC_DIV2K_Reconstructions_q0.zip through q4.zip Five independent archives, 100 images each

Each image archive extracts into its own named directory. Inside it, q0/ through q4/ contain the included operating points. For example:

AFP-GIC_Kodak_Reconstructions_q0-q4/
  q0/kodim01.png
  q1/kodim01.png
  ...
  q4/kodim24.png
  per_image_metrics.csv
  operating_points.csv
  paper_values.csv
  METRIC_PROTOCOL.md
  image_manifest.csv
  README.md

Images preserve their filenames and saved PNG bytes. No resizing or image re-encoding was performed during packaging. The original full image dimensions appear in the per-image CSV files.

Download with Python

Install the download client:

python -m pip install -U huggingface_hub

Download the complete Kodak reconstruction archive and the metrics package:

from huggingface_hub import hf_hub_download

for filename in (
    "AFP-GIC_Kodak_Reconstructions_q0-q4.zip",
    "AFP-GIC_Metrics_and_Documentation.zip",
    "SHA256SUMS.txt",
):
    path = hf_hub_download(
        repo_id="yifeipet/AFP-GIC-Reconstructions-with-Metrics",
        repo_type="dataset",
        filename=filename,
        local_dir="afp_gic_data",
    )
    print(path)

Extract the downloaded ZIP files with your archive manager. For another dataset or operating point, replace the archive filename with one from the table above. These examples download existing results, not model weights or inference code.

Operating points and actual bitrate

Directory quality_index nominal_bpp beta_rate beta_prior
q0 0 0.050 1.921 2.25
q1 1 0.075 1.312 3.50
q2 2 0.100 0.844 3.25
q3 3 0.125 0.422 0.50
q4 4 0.150 0.091 2.25

nominal_bpp identifies the target operating point; it is not a guaranteed bitrate for every image or dataset. Use the bpp column for the measured compression bitrate. The saved PNG is a viewable reconstruction, not the AFP-GIC compressed bitstream. Neither its file size nor the ZIP file size represents the codec's bitrate.

Consolidated CSV files

The metrics archive contains:

File under metrics/ Rows Description
afp_gic_per_image_kodak.csv 120 One row per image and operating point
afp_gic_per_image_clic2020.csv 2,140 One row per image and operating point
afp_gic_per_image_div2k.csv 500 One row per image and operating point
afp_gic_operating_points.csv 15 Dataset-average results retained at paper-source precision
afp_gic_paper_values.csv 15 Corresponding values at the paper's printed precision

Per-image columns start with dataset, then image_name. Identify a reconstruction uniquely by (dataset, image_name, quality_index). Rows are ordered by image name and then operating-point index. Dataset identifiers are kodak, clic2020, and div2k.

Metrics follow the paper's order: bpp, PSNR, SSIM, LPIPS, DISTS, NIQE, followed by the supplementary SNR and MS-SSIM metrics.

Field Meaning Direction
width, height Image dimensions in pixels Not quality metrics
bpp Recorded compressed bits per pixel Interpret together with image quality
PSNR_dB Peak signal-to-noise ratio in dB Higher
SSIM Single-scale structural similarity Higher
LPIPS Learned perceptual image patch similarity Lower
DISTS Deep image structure and texture similarity Lower
NIQE No-reference natural image quality evaluator Lower
SNR_dB Signal-to-noise ratio in dB Higher
MS_SSIM Multi-scale structural similarity Higher
FID Distribution-level metric in the summary CSV only Lower

num_images in the summary CSV is the number of reference images contributing to that dataset/point. Reported means are unweighted means across images: for example, mean PSNR is the mean of individual PSNR values, not PSNR computed from a pooled MSE. Mean bpp is the mean of per-image bpp, not total bits divided by total pixels.

FID is computed at dataset level under the paper's patch protocol. It is not a per-image quantity and must not be obtained by averaging hypothetical per-image FID scores. Kodak FID is blank because it was not reported. Blank values are not zeros.

LPIPS and FID evaluation networks

Metric Feature network Evaluation setup
LPIPS AlexNet, LPIPS v0.1 Saved RGB reconstructions and matching references, scaled to [-1, 1]; full-image evaluation with batch size one
FID FID-specific Inception-v3 from pytorch-fid, 2048-dimensional final pooling features Dataset-level feature distributions from 256 x 256 reference and reconstruction patches

LPIPS evaluation uses lpips.LPIPS(net="alex"), not the VGG variant. Following the HiFiC patch protocol, FID uses a non-overlapping 256 x 256 grid and a second grid offset by 128 pixels in both directions. Only complete patches are retained. FID compares the reference and reconstruction feature distributions separately for each dataset and operating point. See the evaluation implementation.

Comparing with your own method

These reconstructions and metrics allow AFP-GIC to be included as a baseline under matched evaluation protocols without rerunning the pretrained model.

  1. Obtain the original reference images from the benchmark sources cited below. Match images by filename and dimensions.
  2. Evaluate your reconstructions on the same image set, RGB convention, and metric implementation/settings described in METRIC_PROTOCOL.md and the paper.
  3. Compare quality at comparable actual bpp, or compare curves over a common actual-bitrate interval. Matching q indices or nominal targets alone does not establish a rate-matched comparison.
  4. Use the per-image CSVs for individual-image comparisons and the 15-row summary for dataset-level curves. Full-reference metric recomputation requires the original references in addition to these reconstructed images.

PSNR, SSIM, and SNR were evaluated on saved RGB reconstructions. LPIPS was recovered from the same saved reconstructions using LPIPS v0.1 with AlexNet and RGB inputs in [-1, 1], with full-image evaluation at batch size one. Other exported metrics come from archived evaluation records. See METRIC_PROTOCOL.md for provenance and numerical tolerances. Small differences below reported precision can occur between recomputed means and retained paper-source means; values have not been shifted to force agreement.

Original benchmark sources

This release contains model-generated reconstructions of existing benchmarks, not a newly collected set of original photographs. Source-image rights and dataset terms remain with their respective owners. This documentation does not grant a new license to the underlying images. Original reference images should be obtained from their sources.

Integrity and citation

SHA256SUMS.txt lists the upload archive checksums. Inside the metrics archive, IMAGE_MANIFEST.csv maps every PNG to its archive, relative path, byte size, and SHA-256 checksum. PACKAGE_MANIFEST.csv lists the 11 image archives. These small files help check downloads and locate images.

When using these results in research, cite the associated paper. The same data release is also available on IEEE DataPort with dataset DOI 10.21227/z2nw-wp09.

@article{pei2026adaptive,
  author = {Pei, Yifei and Liu, Ying and Ling, Nam},
  title = {Adaptive Fused Prior Transfer for Controllable Generative Image Compression},
  journal = {IEEE Access},
  year = {2026},
  doi = {10.1109/ACCESS.2026.3737467}
}

Contact: Yifei Pei, yifeipet@gmail.com. Authors' affiliation: Santa Clara University.
Downloads last month
59
Total file size:
11.7 GB

Paper for yifeipet/AFP-GIC-Reconstructions-with-Metrics