v1.9.3

Bug fixes

• Fix extension specification error
• Fix TypeError on CPU

Others

• Refactor scripts

v1.9.2

New features

• Add shorter_side and longer_side options

Improvements

• Allow specification of output file name if input is a file

Others

• Refactor scripts

v1.9.1

New features

• Add support for WebP format

Bug fixes

• Fix image reader for transparent PNG
• Fix to open only image files (#14, thanks @ucok66)

v1.9.0

New features

• Check forward path before training
• Add benchmark results

Scale	Noise level 1, YUV420, Scale

Bug fixes

• Fix for benchmark of noise_scale method
• Make -W and -H options unable to use at the same time
• Fix to add noise after subsampling in benchmark

Improvements

• Improve UpResNet10 architecture
• Update UpResNet10 models
• Use chainer.backends
• Change default batch_size and block_size for better performance
• Modify template script of VGG7
• Update UpConv7 noise_scale models (Change initial weights from reference_scale_rgb to anime_style_scale_rgb)

Others

• Reorganize training scripts
• Simplify benchmark options
• Change benchmark graph title

v1.8.1

New features

• Rename ResNet10 to UpResNet10 and improve the architecture
• Add UpResNet10 models and restore old ResNet10 (v1.7.0) models
• Add benchmark script

Improvements

• Performance turning

v1.8.0

Highlights

New ResNet10 models are now available (Experimental)

You can use new ResNet10 models that are about 2x faster than previous one. This feature is experimental, so we recommend using v1.7.0 for stable quality.

New features

• Update ResNet10 architecture
• Add new ResNet10 models
• Add model converters for ONNX and Caffe
• Print elapsed time taken for processing

Improvements

• Optimize loss function
• Modify template training scripts

Others

• Remove .npz files from git history

v1.7.0

New features

• Add reduce_memory_usage option

Bug fixes

• Fix alpha channel processing
• Fix to use init_scope()
• Fix exception handling for Python2

Improvements

• Reduce memory usage
• Change output file name when --model_dir is specified
• Use str.format() instead of %

Others

• Remove unused code

v1.6.0

New features

• Add width and height options
• Add support for transparent PNG
• Add support for denoising with UpConv7 model
• Parse tiff images

Bug fixes

• Fix a bug that images cannot be scaled more than 2 times
• Fix failing to specify output directory
• Make sure loading files containing multi-byte characters
• Fix a bug that denoising is performed multiple times

Improvements

• Simplify some options
• Change behavior of --output

Others

• Remove ResUpConv10 models
• Rename --scale_factor to --scale_ratio

v1.5.0

New features

• Show elapsed time per inner epoch
• Add ResNet10 RGB models
• Add ResNet10 Y-only models

Bug fixes

• Fix memory allocation error

Improvements

• Chainer v2 compat
• Change six.print_ to print

Others

• Remove computational graph writing tools
• Remove reference_scale models

v1.4.0

New features

• Add ResUpConv10 RGB models

Bug fixes

• Hotfix for ResUpConv10 architecture