v5.1.0

• Add num_batches parameter.
• Add AnimeJaNai_HD_V3Sharp1 and AnimeJaNai_SD_V1beta34 models from https://github.com/the-database/mpv-upscale-2x_animejanai.
• Switch to dynamo IR for compilation in TorchTRT. A few parameters are added and INT8 quantization is removed.
• Improve performance by using separate streams and non_blocking for moving tensors between CPU and GPU. Now num_streams arg has negligible influence on performance compared to num_batches arg.
• Bump PyTorch to 2.6.0.dev.
• Bump Torch-TensorRT to 2.6.0.dev.

v5.0.0

• Add 2x_AnimeJaNai HD V3 models from https://github.com/the-database/mpv-upscale-2x_animejanai.
• Add AniScale 2, OpenProteus and Ani4K v2 models from https://github.com/Sirosky/Upscale-Hub.
• Change default model to AnimeJaNai_HD_V3_UltraCompact_2x.
• Remove nvfuser and cuda_graphs.
• Add support for TensorRT INT8 mode using Post Training Quantization (PTQ).
• Bump PyTorch to 2.4.0.dev.
• Bump TensorRT to 10.0.1.
• Bump Torch-TensorRT to 2.4.0.dev.
• Bump VapourSynth to R66.

v4.1.0

• Place model files under models directory for cleanness.
• Don't globally set default floating point tensor type when the input is of RGBH format.
• Change function name to lowercase.
• Bump required version of dependencies.

v4.0.1

• Switch to PyTorch again for inference. A few parameters are added and some parameters are removed.
• Add official ESRGAN x4 model and realesr-general-x4v3 model.

See Discussions for benchmarks of some models.

v3.1.0

• Replace RealESRGANv2-animevideo model with realesr-animevideov3 model.
• Change the default of model to 3.

v3.0.0

• Switch to ONNX Runtime for inferencing.
• Rename model_type parameter to model.
• Rename tile_x, tile_y parameters to tile_w, tile_h.
• Remove pre_pad parameter.

v2.0.0

• Add a check to make sure model files have been downloaded.
• Only VS API4 is supported now.
• Add support for RealESRGANv2-animevideo models.
• Remove scale and anime parameters.
• Add model_type parameter.

v1.2.0

• Change tile parameter to tile_x and tile_y.
• Add support for VS API4.

v1.1.0

• Add model optimized for anime.
• Rename half parameter to fp16 for intuition.

v1.0.0

• Initial release.