I actually tried, couldn't do it. Some version of numpy in their requirements.txt doesn't build with the Python on the my system.

Sure, I could probably install another Python with conda, but it's not my idea of fun. The app here actually worked fine from the first try, without taking gigabytes of disk space.

OK. That makes some sense. I tend to assume everyone with a CUDA-capable GPU is already using conda. But of course that's not true.

I followed the instructions in their README, they suggest conda, but don't tell you to use it.

My GPU isn't even CUDA-capable, but the previous version of Upscayl worked fine, possibly on the CPU. That might have changed, I don't know.

In any case, even with conda installing everything would have been a pain to someone not really accustomed to the Python ecosystem. I don't care too much about the GUI, but a working AppImage is really nice.

What GPU do you have? The Upscayl README says in big letters:

"NOTE: Upscayl does not work without a GPU, sorry. You'll need a Vulkan compatible GPU to upscale images. CPU or iGPU won't work."

Sure, perhaps a previous version may have worked without a GPU, but this seems surprising given that it uses a GPU-only library for all the heavy lifting.

The original python version uses pytorch/CUDA and requires nvidia. The Vulkan version doesn't.

Yeah, well, both versions work on my system. I don't know how to check if I'm getting any kind of acceleration, probably not on AMD.

EDIT: I think it's actually using Vulkan Compute on RADV.

Well you can use the ncnn version that is the backend for this program. https://github.com/xinntao/Real-ESRGAN-ncnn-vulkan

I've always been confused with Conda, as it seems like overkill compared to just pyenv.

The fact that its not up to date on pypi and looking at their "getting started" is tens of pages of Windows bugs has always stopped me using it.

Yeah you know Python's software distribution problems are bad when someone makes an entire parallel packaging universe.

I do machine learning and use Python a lot, and even then I'd sometimes prefer something packaged like this. Dependencies are Python's hell and I can't usually know in advance if a repo will be a 1-minute install or a 15-minute install.

Visual feedback. Visual interface. The GUI fad is still around.

Right, but if a GUI is the only requirement, it might be simpler to use gradio.

It only takes a few lines (pasted from https://huggingface.co/spaces/akhaliq/Real-ESRGAN):

  gr.Interface(
      inference, 
      [gr.inputs.Image(type="pil", label="Input"),gr.inputs.Radio(["base","anime"], type="value", default="base", label="model type")], 
      gr.outputs.Image(type="file", label="Output"),
    title=title,
    description=description,
    article=article,
    examples=[
    ['bear.jpg','base'],
    ['anime.png','anime']
    ]).launch()