xjpeg 0.1.1

Extending JPEG with neural networks (X-JPEG): Image-adaptive quantization with neural networks for JPEG

X-JPEG

Extending JPEG with neural networks (X-JPEG): Image-adaptive quantization with neural networks for JPEG

X-JPEG analyzes an image, predicts three 8×8 quantization tables for Y, Cb, and Cr, and passes them to a production JPEG encoder. The output is a normal JPEG: browsers, operating systems, and existing decoders need no plugin or neural network.

image ──► compact CNN + full-image DCT statistics ──► adaptive Y/Cb/Cr DQTs
                                                        │
                                                        ▼
                                              MozJPEG ──► .jpg

The current implementation and release weights were developed by Migel Tissera in 2026 and are released by Trinity Cloud under the MIT License.

Install

pip install xjpeg

Official platform wheels bundle MozJPEG. On an unsupported platform, X-JPEG automatically uses a verified system MozJPEG (XJPEG_MOZJPEG=/path/to/cjpeg) or falls back to Pillow/libjpeg.

Compress

# Complete-file bits per source pixel. This is conventional image-codec bpp.
xjpeg photo.png --target-bpp 0.5

# Search for the smallest file that reaches the requested RGB MS-SSIM.
xjpeg photo.png --target-msssim 0.98

# Directly control the learned tables (>1 means coarser/smaller).
xjpeg photo.png --scale 1.5

Progressive JPEG and a shared emitted chroma DQT are the release defaults. The network still predicts independent Cb and Cr tables. Use --three-dqt to emit all three tables or --sequential to disable progressive encoding.

from xjpeg import XJPEG

codec = XJPEG()  # auto-selects bundled/system MozJPEG
result = codec.compress(
    "photo.png",
    output="photo.jpg",
    target_bpp=0.5,
)

print(result.bpp, result.msssim, result.backend)
print(result.luma_table)

Low-rate results

Deterministic sample of 100 held-out native-resolution COCO val2017 images, seed 20260721. Rates include the complete emitted file. Every JPEG row used the same pinned MozJPEG build and settings; WebP used method 6. The release model predicts three tables and the default deployment averages Cb/Cr only at the encoding boundary.

0.50 bpp target	Actual bpp	RGB MS-SSIM ↑	Y MS-SSIM ↑	PSNR ↑
X-JPEG, default 2-DQT emission	0.49975	0.957182	0.967685	27.045 dB
X-JPEG, native 3-DQT emission	0.49917	0.956934	0.967261	27.015 dB
Standard Annex-K tables + MozJPEG	0.50097	0.953630	0.972323	26.958 dB
WebP method 6	0.50015	0.956685	0.972758	29.164 dB

0.25 bpp target	Actual bpp	RGB MS-SSIM ↑	Y MS-SSIM ↑	PSNR ↑
X-JPEG, default 2-DQT emission	0.25111	0.917638	0.932246	24.899 dB
X-JPEG, native 3-DQT emission	0.25002	0.916420	0.930914	24.853 dB
Standard Annex-K tables + MozJPEG	0.25000	0.912369	0.938567	25.009 dB
WebP method 6	0.24996	0.922521	0.942144	26.609 dB

The result is metric-specific. X-JPEG improves RGB MS-SSIM over the declared standard-table JPEG control at these operating points; WebP retains a clear advantage at 0.25 bpp and on luma MS-SSIM/PSNR. The 0.50-bpp X-JPEG/WebP mean difference is too small to claim a general win. See docs/BENCHMARKS.md for the protocol and limitations.

Train

pip install "xjpeg[train]"

python -m xjpeg.train \
  --data /path/to/train2017 \
  --val /path/to/kodak \
  --out runs/experiment \
  --architecture compact \
  --num-tables 3 \
  --dct-stats \
  --gdn-reparam \
  --native-crops \
  --lambda-r 0.05 \
  --lambda-luma 0.25

Training combines differentiable JPEG distortion, a learned coefficient-rate proxy, soft table entropy, and an optional autoencoder reconstruction loss. Release evaluation always uses real encoded files and complete-file byte counts. The full objective and validation contract are documented in docs/METHODOLOGY.md.

Model and releases

• Source and issues: https://github.com/trinity-cloud/x-jpeg
• Model weights: https://huggingface.co/migtissera/x-jpeg
• Python package: https://pypi.org/project/xjpeg/
• Model card: MODEL_CARD.md

Limitations

• Trained and evaluated primarily on natural photographs.
• Not validated for medical, scientific, text-heavy, or adversarial imagery.
• Encoding is slower than libjpeg because it adds neural inference and MozJPEG trellis optimization; decoding speed is unchanged.
• EXIF/ICC metadata is not preserved in version 0.1.0.
• --target-bpp and --target-msssim search per image and therefore cost multiple real encodes.

License

X-JPEG code and weights are MIT licensed. Bundled MozJPEG is distributed under its compatible upstream licenses; see THIRD_PARTY_NOTICES.md.