Real-ESRGAN upscale models v1

Real-ESRGAN upscale models

Pre-packaged model weights for vhs_upscale.sh / vhs_upscale_bw.sh /
vhs_upscale_anime.sh (and vhs-gui's Upscale panel), so you don't have to
hunt these down individually.

Assets

File	Contents	Extract to
realesrgan-models-ncnn.zip (236 MB)	.bin/.param pairs — Vulkan backend	~/opt/realesrgan-ncnn/models/
realesrgan-models-pytorch.tar.gz (336 MB)	.pth checkpoints — ROCm backend	~/opt/realesrgan-rocm/models/

You generally only need one archive, matching whichever UPSCALE_BACKEND
you use. Both include the same model set in their respective format.

What's included

• Official Real-ESRGAN models (BSD-3-Clause, unmodified upstream
  weights from xinntao/Real-ESRGAN):
  realesrgan-x4plus, realesrgan-x4plus-anime, realesr-animevideov3
  (x2/x3/x4), RealESRGAN_x2plus, RealESRGAN_x4plus_anime_6B,
  RealESRGAN_x4plus
• Community VHS-tuned models (CC-BY-NC-SA-4.0, third-party, not
  covered by this repo's GPLv3 — see MODEL-CREDITS.md in each archive
  for full attribution): 2x_VHS-upscale-and-denoise_Film (Itewreed),
  1x-VHS-Sharpen (RTX 2080 ti hoarder), ToonVHS-1x (Redslam), all
  sourced from OpenModelDB

Install

mkdir -p ~/opt/realesrgan-ncnn/models
unzip realesrgan-models-ncnn.zip -d ~/opt/realesrgan-ncnn/models

mkdir -p ~/opt/realesrgan-rocm/models
tar xzf realesrgan-models-pytorch.tar.gz -C ~/opt/realesrgan-rocm/models

Vulkan backend note: realesrgan-ncnn-vulkan hardcodes its network
architecture by matching the model name against a fixed family list
(realesrgan-x4plus, realesrgan-x4plus-anime, realesrnet-x4plus,
realesr-animevideov3*) — the community VHS models are ROCm-only;
selecting them under UPSCALE_BACKEND=vulkan fails fast with a clear error
(see the main README's "Choosing a model" table).