Advanced RealESRGAN Settings (Manual Tile Size Override for High-End GPUs)

Hi,

First of all, thank you for developing Video2X. It is one of the best and most user-friendly AI video upscalers available.

While experimenting with Video2X 6.4.0, I noticed that the RealESRGAN backend always relies on the automatic tile size selection from librealesrgan_ncnn_vulkan.

After looking through the source code, I found the following logic in filter_realesrgan.cpp:

uint32_t heap_budget = ncnn::get_gpu_device(gpuid_)->get_heap_budget();

if (heap_budget > 1900) realesrgan_->tilesize = 200; else if (heap_budget > 550) realesrgan_->tilesize = 100; else if (heap_budget > 190) realesrgan_->tilesize = 64; else realesrgan_->tilesize = 32;

This works well for compatibility and I completely understand why these automatic values exist.

However, there is currently no way for advanced users to override the tile size and experiment with different values on modern high-end GPUs.

Why this would be useful

I'm currently running Video2X on an AMD Radeon RX 9070 XT (16 GB VRAM).

With realesr-animevideov3 (2× upscale), I currently get around 17 FPS, which is already excellent.

However:

GPU memory usage remains relatively low. The GPU is not always fully utilized. There is no possibility to determine whether a larger tile size could further improve throughput.

I'm not requesting a change to the default behavior.

Instead, I'd like to suggest an optional Advanced Settings section, for example:

☑ Automatic tile size (default) ☐ Custom tile size

If "Custom" is selected, the user could specify values such as 256, 512, 768, 1024 or higher.

This would allow experienced users to optimize performance for their own hardware while keeping the current safe defaults for everyone else.

Possible future advanced options

If an "Advanced" section is ever added, it could also become the place for other backend-specific expert settings where applicable, while keeping the main interface clean and beginner-friendly.

I'd also be happy to test experimental builds and provide benchmark results on modern AMD hardware if that would be useful.

Thank you again for your work!