v1.6.2

Minor release to improve the chroma models.

Changed:

• Both models: The degradation model has changed from using vanilla JPEG noise to using custom 4x4 DCT blocks with a monotonically increasing radial base quantisation table. This change was done to better match compression artifacts from codecs like AVC, HEVC and AV1.
• ArtCNN_R8F64_Chroma: Now trained with minor compression noise to improve general usefulness.
• ArtCNN_R8F64_Chroma_DN: Tuned to be slightly more conservative in its denoising capabilities.

v1.6.1

Minor release to improve backwards compatibility with existing scripts.

Changed:

• Inputs and outputs in the ONNX models have been renamed to input and output, respectively. This should solve compatibility issues with scripts that have the input name hardcoded to input. Previously, jax2onnx had the inputs named in_0_nchw, which is fine but needlessly verbose.

v1.6.0

Minor release to improve the ONNX exports.

Changed:

• All ONNX models are now exported using jax2onnx. The new exporter is better in many ways, so the resulting ONNX models look a little cleaner and are better supported by Netron. For now, the models themselves have stayed the same, the weights have been simply transplanted from the Keras models to the corresponding Flax NNX models.

My aim is to have all future models and experiments done in a pure JAX environment, without having to rely on Keras or TensorFlow. The remaining hurdle was the lack of a reliable ONNX exporter, but jax2onnx is mature enough now.

v1.5.2

Minor release to adjust ArtCNN_R8F64_Chroma_DN.

Changed:

• ArtCNN_R8F64_Chroma_DN: Denoising strength increased back to match the other DN and the JPEG models. This provides better reconstruction quality overall and does look better when fine detail isn't lost. The previous change was aimed at increasing the amount of fine detail retained, at the expense of reducing denoising strength which looked worse on a variety of shows.

v1.5.1

Minor release to adjust the Chroma models.

Changed:

• ArtCNN_R8F64_Chroma: Degradation model simplified back to being purely a chroma upscaling model. Compression noise had been previously added to help prevent artifacts with noisy content, but a proper variant aimed entirely at noisy content exists now. The issues that had been fixed by training with noisy inputs seem to have stayed fixed after fine-tuning with clean data.
• ArtCNN_R8F64_Chroma_DN: Denoising strength reduced to better preserve fine-details. This new middle-ground offers a better compromise on most tested sources.

The goal is to reduce the amount of redundant options and set clear use-cases. Previously, there wasn't much of a difference between Chroma_DN and the JPEG420 model. The former has been adjusted to be a middle ground between the 2 extremes, so there are 3 distinct levels of denoising strength effectively being offered now.

v1.5.0

This marks the first release in 2026!

New:

• ArtCNN_R8F64_Chroma_DN: Trained to upscale chroma from noisy content. This model cleans compression artifacts very aggressively.

Changed:

• ArtCNN_R8F64_Chroma: Faint compression artifacts added to the degradation model to prevent the model from going haywire when some faint compression noise is present in the input.

Removed:

• ArtCNN_R16F96_Chroma: Didn't seem to provide enough of an improvement over the smaller version to justify its existence. Can be brought back in the future if there's any demand.

New Experiments:

• ArtCNN_R8F64_JPEG444_YCbCr: Experimental BT.709 YCbCr version of the JPEG444 model.
• ArtCNN_R8F64_JPEG420_YCbCr: Experimental BT.709 YCbCr version of the JPEG420 model.

v1.4.2

Minor release to make the chroma models less likely to hallucinate when used on lossy/noisy content. Further improvements might be done in the future to make them more robust in general.

All other models have stayed the same.

v1.4.1

Minor release to increase denoising strength on both the DS and DN variants. All other models have stayed the same.

This release also has a revamped ONNX Inferencer, making it easier to use the ONNX models on images. The new inferencer is capable of doing batch processing, and it also supports geometric self-ensembling on all tasks (upscaling, denoising, etc).

v1.4.0

This release is bigger than usual, so I'll mark it as 1.4.0!

Many changes have been made since the last release. These vary from new models to improvements to the existing ones.

New models:

• ArtCNN_R8F64_JPEG444: 1x RGB model meant to clean JPEG artifacts.
• ArtCNN_R8F64_JPEG420: 1x RGB model meant to clean JPEG artifacts. Also fixes chroma subsampling.
• ArtCNN_C4F32_DN: 2x Luma model that also denoises. Works well on noisy sources when you do not want any sharpening.
• ArtCNN_C4F16_DN: Same as above.

Updated models:

• ArtCNN_R16F96: Works better with higher bit-depth sources now.
• ArtCNN_R8F64: Same as above.
• ArtCNN_C4F32_DS: Stronger denoising, cleaner look overall.
• ArtCNN_C4F16_DS: Same as above.

This is the first release in a while with changes to the C models, so any feedback is appreciated.

v1.3.1

This marks the first minor release in the 1.3.x series!

Changes:

• ArtCNN_R16F96_Chroma: Self-explanatory, just the R16F96 version of ArtCNN Chroma. This provides unprecedented chroma reconstruction quality. Easily above my expectations.
• ArtCNN_R8F64_Chroma: The bigger version was performing so much better it prompted me to revisit this model and to train it for a little longer on FP32. Quality is a bit improved but it's still roughly in the same ballpark as before.

The other models have all stayed the same.