/Upscale-Hub

A repository collecting image and video upscaling resources as well as my own super resolution models.

OtherNOASSERTION

Upscale Hub: a repository of resources for upscaling and neural network training, as well as a collection of the models I've trained.


animated

Comparison image for AniScale 2 (WIP).

✨ How to Upscale

📈 How to Train Your Own Model

If you're looking for a guide on how to train your own model, check out the following:

⏬ Released Models

Ratings included for each model are meant to serve as a quick point of reference to see how the models differ.

DVD/HD/FHD Sources - AniScale 2 [Gen 6, OmniSR, Compact, etc.]

[Gen 6, OmniSR, Compact, etc.] The successor and replacement to the original AniScale, and a substantial upgrade in nearly every respect. AniScale 2 is a versatile and faithful anime model trained for use on a variety of post ~2000 sources. Superb blur and depth of field handling, thorough WEB and DVD compression repair, and pleasing line art refinement are the hallmarks of AniScale 2.

  • While AniScale 2 is trained first and foremost as an OmniSR model, AniScale 2 is also intended to be a platform to explore multiple SISR archs. For starters, I've trained OmniSR and Compact versions, but more will come.

  • AniScale 2 also comes with a "refiner" model, creatively named AniScale 2 Refiner (AS2R). AS2R is a 1x Compact model (for maximum speed) trained to supplement and increase the versatility of AniScale 2, without making the base model excessively aggressive. The Refiner is focused on providing light sharpening, fixing line art and line thinning, depending on whether you run the model before or after upscaling with AniScale 2.

  • For more information, please visit the Github wiki page.

  • Download

    • If you're unsure which version to start with, I recommend trying the OmniSR version first. More details on the Wiki page linked above.
  • Image Comparisons:

  • Detail Retention: ⭐⭐⭐⭐

  • Compression Cleanup: ⭐⭐⭐⭐⭐

HD/FHD Live Action TV and Cinema - Open Proteus [Compact]

[Compact] Yes... not an anime model for a change! Topaz Video AI is generally considered amongst the industry leaders of in upscaling movies and TV, however it sits behind a significant paywall. Open Proteus is a free alternative to Topaz's Proteus AI, and intended to upscale high-quality and low-noise HD sources to 4K. While I won't claim this is superior to Topaz Proteus by most metrics (which would be a near-impossible feat on open-source architectures anyway), Open Proteus produces output which is highly faithful, without causing oversharpening or upscaling artifacts.

HD/FHD Sources - Ani4K [Gen 5, Compact]

[Gen 5, Compact] As the name suggests, Ani4K is trained to create natural-looking 2K/4K upscales from 720p/1080p sources. In particular, the model has superb detail retention.

  • Download
  • Image Comparisons:
  • Detail Retention: ⭐⭐⭐⭐⭐
  • Compression Cleanup: ⭐⭐(will deal with compression in HD/FHD BD and WEB sources without a hitch, but may be less suited for DVDs-- I'll be releasing separate models for DVDs)

Legacy Model - AniScaleV1 [Gen 1, Compact]

My very first publicly released model. I consider it obsolete at this point, but some people might prefer the sharper look, so I'm releasing it here as well. As a first generation model, I trained AniScale to serve as a general-purpose model for all sources. Unfortunately, it simply isn't possible to have a single model do everything well-- subsequent generations of models are more specialized, and will do better for the sources they specialize in. While the model is great with details and dealing with a variety of video artifacts, it has a tendency to oversharpen some sources and background details. It doesn't handle blur at all either, though that's a problem common with many other anime models.

  • Download
  • Image Comparison
  • Detail Retention: ⭐⭐⭐⭐
  • Compression Cleanup: ⭐⭐⭐⭐
  • Blur / DOF Effect Retention: ⭐ (oversharpens out-of-focus objects)
  • Upscaling Artifacts: ⭐⭐⭐ (slight color shifting on some objects)

🤝 Tools Utilized and Acknowledgements

  • chaiNNer by Joey and contributors: The backbone of any of my bulk image processing and upscaling process. These models would not be possible without ChaiNNer.
  • ImgAlign by sonic41592: A fantastic little tool which can automate one of those tedious parts of image pair dataset preparation.
  • Image Pearer: A tool to automate the image pairing process by yours truly. Allows automation of the other most tedious part of the image pair dataset preparation process.
  • OpenModelDB: A fantastic resource for all things upscaling.
  • neosr by muslll: The hottest new training platform!
  • TraiNNer Redux and Forks by Joey, muslll and other contributors. While I've largely moved over to neosr, I've spent a LOT of time on Redux.
  • Dataset Destroyer by Kim2091: A crucial part of the dataset preparation process that makes image degradation oh so easy!
  • The developers and contributors to neural network models such as ESRGAN, SRVGGNetCompact, OmniSR, SwinIR and more for developing the networks on which these models are based.
  • Simple Image Compare: Another tool by yours truly. A minimalistic image comparison tool which I developed because ICAT sucks.
  • And last but certainly not least, the Enhance Everything Discord community for their invaluable support and guidance. I also blame them for my neural network training addiction.