Skip to content

Performs worse than bicubic on photographic content for SISR #9

@catid

Description

@catid

I evaluated this software against the state of the art for learned upsamplers at 60 FPS, targeting both desktop GPUs and Intel iGPUs, and 1920x1080 output resolution. The results are detailed here: https://catid.io/posts/tiny_sr/

Perhaps I'm not using the software correctly. To experiment, this is my methodology:

It compiles cleanly with Visual Studio 2022 on Windows. If you build the DX12 sample, add images under NVIDIAImageScaling\samples\bin\DX12\media\images, run the DX12 sample application, and then select 50% scale and one of the images from the drop-down, you can press the S key to have it save the image to NVIDIAImageScaling\samples\bin\DX12\dump.png.

Please advise if there is a better way to evaluate the performance of the upsampler on a given PNG image.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions