I asked ChatGPT plus
https://chatgpt.com/share/6b00382b…f8-47e253a8cff8
And here's where the AI explanation becomes questionable: NGU Anti-Alias scores even worse than NGU Sharp. Since Anti-Alias isn't primarily intended to sharpen aggressively, simply blaming sharpening doesn't adequately explain these results.
Other possible explanations include the test material, the NGU quality settings, the resizing pipeline, and how the original reference was generated.
What I would want to know about the developer's benchmark
Before concluding that NGU is inferior, I'd want to establish:
- Was NGU tested at Low, Medium, High or Very High quality?
- Were all algorithms operating on identically prepared source frames?
- Were additional sharpening or processing stages disabled?
- Were SSIM, VMAF or subjective viewing comparisons performed alongside PSNR?
These details are not specified in the screenshot.
My conclusion
The developer's observation that FSRCNNX performs consistently well according to PSNR is supported by the chart.
But the statement that NGU isn't convincing should be qualified: NGU isn't convincing in this particular pixel-fidelity benchmark. That is not the same as proving it delivers worse perceived image quality.
Scientifi.c research supports this distinction between numerical reconstruction accuracy and perceived detail.
One particularly useful follow-up for the LibreELEC CB1 developer would be to compare NGU, FSRCNNX and Lanczos using SSIM or VMAF alongside side-by-side images. That would help determine whether NGU is genuinely losing important detail or simply being penalised for reconstructing it differently.