Add Upscale Face / Remove Background nodes in NCNN (for older AMD GPU)

[Disonantemus]

Motivation

• My GPU is AMD and very old (RTX 580) and I don't have the money 2 upgrade.
• I can use NCNN models that do use GPU, and are A LOT faster than CPU only.
• PyTorch and ONNX sections work, but with CPU only and VERY SLOW.
• ROCm only support a few newer/expensive AMD GPU.

Description

• Add "Upscale Face" node from PyTorch section into NCNN (very useful).
• Add "Remove Background" node from ONNX section into NCNN (optional).

Alternatives

I don't know of any alternative, maybe:

• Allow subType: "FaceSR" when converting to NCNN model (right know shows an error).
• Allow to upscale with subType:"FaceSR" with "Upscale Image" node (NCNN); and do later the post processing blending with both images (Face+Background) in an editor (like GIMP).

---

I'm very happy and grateful with the things that you can do in ChaiNNer (without dependency hell or knowing python).

[joeyballentine]

The face models cannot be converted to NCNN as far as I'm aware. Not sure about the background removal ones though

[Disonantemus]

• Is there an easy way to use those models besides ChaiNNer (or maybe ComfyUI)?
• Maybe, as a regular upscaler? and do the blending later in an editor.

---

I found those same models in ONNX format within ChaiNNer, but there is no "Upscale Face" in the ONNX section.

I did found today RestoreFormerPlusPlus in .ckpt that didn't work in ChaiNNer. Later I found an ONNX version, but not a .pth, then because ONNX section doesn't have "Upscale Face", then, I can't test it.

But it doesn't help either, because my GPU doesn't work in ONNX mode, only CPU.

[Efenstor]

For face upscale with NCNN there is a half-abandoned project: https://github.com/onuralpszr/GFPGAN-ncnn-vulkan. A couple of years ago I tried it and it worked. It uses a somewhat outdated model GFPGANCleanv1-NoCE-C2. Still it cannot be loaded into chaiNNer as is.

[Disonantemus]

For face upscale with NCNN there is a half-abandoned project: https://github.com/onuralpszr/GFPGAN-ncnn-vulkan. A couple of years ago I tried it and it worked. It uses a somewhat outdated model GFPGANCleanv1-NoCE-C2. Still it cannot be loaded into chaiNNer as is.

I was able to compile it, as it is, has some quirks:

• No help at all.
• No custom parameters (as other ncnn upscalers), like: scaling, output file, model folder, denoise, etc.
• Only 2x.
• Only change faces, not clothes or background (from code, should use real_esrgan 4 that).
• No model folder.
• Fixed output filename: result.png

Is better than nothing.

---

There is a fork, but got worst experience:

• 512x512 output (square aspect ratio).
• Non realistic results.
• Has a lot of strange commits "auto" (1607 commits ahead) doing similar things like:

- import os,sys,aifc
- l=3
+ import os,sys,errno
+ d=9