Skip to content

Release [1.6.0] - DirectML Upscaling & Interpolation
Compare
Choose a tag to compare
@NevermindNilas NevermindNilas released this 27 Apr 22:00
· 1698 commits to main since this release
v1.6.0
9cdabf6

New Features

  • DirectML Backend: A new backend for upscaling and interpolation. Personal tests show approximately 20% faster inference speeds over NCNN in interpolation and anywhere from 2x-4x for upscaling over NCNN. In some cases, it seems to match or even surpass CUDA.
  • Offline Mode: A new --offline flag allows the user to download all available models in one go, making the script fully offline (with the exception of YT-DLP and Depth Maps).
  • Data Sharing: A new --consent flag allows users to opt-in to share select data from log.txt to a server, mainly meant for improving the script in edge case situations.
  • Rife 4.15-lite: This version has been added.
  • New Models: The following models have been added: "compact-directml", "ultracompact-directml", "superultracompact-directml", "span-directml", "cugan-directml", "shufflecugan-directml"

Improvements

  • Performance: This release includes several performance improvements, primarily for interpolation with the introduction of a slightly better caching functionality. Interpolation speeds for Rife-Cuda should have increased by roughly 5% over previous releases.
  • Code Structure: The structure of the code has been slightly adjusted.
  • Logging and Printing: Logging and printing functionalities have been improved.
  • Download Functionality: Switched from wget to requests + tqdm to bypass certain errors some users were running into.
  • File Size: The total file size of TAS has been reduced from 1.5.0 to 1.6.0 by approximately 300MB.
  • GUI: The AE GUI now saves the previous set settings across launches. The backend and layout of the AE GUI have also been improved.
  • Input Handling: There is likely better handling of edge case inputs with uncommon characters.
  • Models: New Cugan Model.

Removed

  • NCNN Backend: The NCNN backend has been removed due to its slower speed compared to DirectML and its large space requirement.
  • Removed Models: Removed Cugan's default models in favor of phhofm trained model for better results.
  • Removed flags: Removed --cugan_kind for the above reason.

Fixed

Benchmarks

13700k Undervolted + Overclocked, 3090 Overclocked, 32GB DDR5 6800Mhz, TAS 1.5.0/1.6.0, Nilas.

TAS 1.6.0
tables

TAS 1.5.0
tables

red2_00000
Ty to Renarchi for the thumbnail.

Assets 3
Loading