X-JPEG 0.1

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

X-JPEG predicts image-adaptive JPEG quantization tables. The encoder produces three 8×8 tables for Y, Cb, and Cr; those tables are passed to MozJPEG or libjpeg to produce an ordinary, standards-compliant .jpg. No neural network or custom software is required to decode the output.

Developed by Migel Tissera / Trinity Cloud.

Artifact

File	Purpose	SHA-256
model.safetensors	Encoder-only weights	823a305b9cdf6d16e2296644d88a343dfc68621fa8e3d4b3c0125772c50b92e8
config.json	Architecture and preprocessing contract	f4766ca0045dc256645054da19dda60401e9f05d1438f1ef986e77bfd0cf5e87
benchmark.json	Sanitized low-rate aggregate results	41ffb4739cb58e5687a8a200b27b1c5cfd016aae5d823f79a6cae5d22cc791a6

The source PyTorch checkpoint was converted with weights_only=True and prediction parity was verified exactly on a deterministic non-square RGB input. This repository contains no pickle checkpoint or training optimizer state.

Use

The PyPI package already contains these weights:

pip install xjpeg
xjpeg photo.png --target-bpp 0.5

To load this Hub snapshot explicitly:

from pathlib import Path
from huggingface_hub import snapshot_download
from xjpeg import XJPEG

snapshot = Path(snapshot_download(
    "migtissera/x-jpeg",
    allow_patterns=["model.safetensors", "config.json"],
))
codec = XJPEG(snapshot / "model.safetensors")
result = codec.compress("photo.png", output="photo.jpg", target_bpp=0.5)
print(result.bpp, result.msssim, result.backend)

Source, training code, and methodology: https://github.com/trinity-cloud/x-jpeg

Model architecture

• 256×256 RGB thumbnail input.
• Four 5×5 stride-2 convolution + reparameterized GDN blocks.
• Final stride-2 head producing a 3×8×8 table bottleneck.
• Frequency-aligned 3×8×8 conditioning from full-resolution DCT-band energy.
• Integer tables mapped to [1, 255] and stored in JPEG DQT segments.
• Three predicted tables; the package defaults to averaging Cb/Cr only at emission time to reduce low-rate header overhead.

The model has 19 tensors in its encoder artifact. The training-only mirror decoder, table entropy model, coefficient rate model, and optimizer state are not included.

Training

The model was trained on COCO train2017. Its differentiable 4:2:0 JPEG objective combines RGB MS-SSIM, a 0.25-weight luma MS-SSIM guard, and a learned DCT-symbol rate proxy with rate weight 0.05. Release measurements use actual encoded files rather than the rate proxy.

Low-rate evaluation

100 deterministic held-out native-resolution COCO val2017 images, seed 20260721, complete-file conventional bpp, pinned MozJPEG pipeline:

Method	bpp	RGB MS-SSIM ↑	Y MS-SSIM ↑	PSNR ↑
X-JPEG default	0.49975	0.957182	0.967685	27.045 dB
Annex-K + same MozJPEG	0.50097	0.953630	0.972323	26.958 dB
WebP method 6	0.50015	0.956685	0.972758	29.164 dB
X-JPEG default	0.25111	0.917638	0.932246	24.899 dB
Annex-K + same MozJPEG	0.25000	0.912369	0.938567	25.009 dB
WebP method 6	0.24996	0.922521	0.942144	26.609 dB

These results support an RGB MS-SSIM improvement over the declared standard-table JPEG control. They do not establish general superiority over WebP; WebP is clearly ahead at 0.25 bpp and in luma/PSNR.

Intended use and limitations

Intended for lossy compression of natural photographs and research on adaptive JPEG quantization. It is not an archival or forensic-preservation codec.

• Other image domains, including medical and scientific imagery, are not established.
• Version 0.1.0 does not preserve EXIF or ICC metadata.
• --target-bpp performs multiple real encodes per image.
• Encoding is slower than libjpeg; decoding remains ordinary JPEG decoding.
• The 100-image low-rate evaluation should be expanded before making strong frontier claims.

License

The X-JPEG code and model weights are released under the MIT License. MozJPEG is not included in this model repository; platform wheels may bundle it under its upstream license notices.