Local AI Image Upscaler — AMD GPU (Vulkan)

Local upscaler for AI-generated images, aimed at high-quality print output. Uses Real-ESRGAN NCNN Vulkan — no CUDA, no cloud, native AMD GPU support on Windows.

Tested on AMD Radeon RX 9060 XT, Windows 11.

---

Quickstart

pip install -r requirements.txt

# Photorealistic / AI photo-realistic (best quality)
python upscale.py my_image.png

# AI art, illustration, anime
python upscale.py my_image.png --model anime

# Fast preview
python upscale.py my_image.png --model fast

# 2x instead of 4x
python upscale.py my_image.png --scale 2

# Process a whole folder
python upscale.py images/ --model photo

# Custom output folder
python upscale.py my_image.png --output ~/Pictures/upscaled

The NCNN binary and models (~43 MB) are downloaded automatically on first run. By default, results land in output/.

---

Benchmark — AMD RX 9060 XT

Input: 768×1364px → Output: 3072×5456px (4× scale, >4K on the long edge)

Model	Time	Resolution	PNG size	Recommended use
photo (x4plus)	9.6s	3072×5456px	19.6 MB	Photorealistic images, AI portraits
anime (x4plus-anime)	3.8s	3072×5456px	17.7 MB	AI art, illustration, cartoon
fast (animevideo)	1.6s	3072×5456px	21.4 MB	Previews, batch processing

All models autodetect the GPU via Vulkan (fp16 enabled).

Print DPI guide

Output resolution	300 DPI (photo quality)	150 DPI (wall poster)
3072×5456px (4×)	~10" × 18"	~20" × 36"
1536×2728px (2×)	~5" × 9"	~10" × 18"

---

Available models

File in bin/models/	--model flag	Scale	Weight
realesrgan-x4plus	photo	4×	32 MB
realesrgan-x4plus-anime	anime	4×	8.6 MB
realesr-animevideov3-x2/x3/x4	fast	2/3/4×	1.2 MB

---

Manual install (optional)

The script auto-downloads the binary + models on first use. To do it manually:

upscale.py          <- main script (single Python file)
requirements.txt    <- only pillow
bin/
  realesrgan-ncnn-vulkan.exe   <- NCNN binary
  vcomp140.dll                 <- Visual C++ runtime
  models/
    realesrgan-x4plus.bin/.param         <- photo model (32 MB)
    realesrgan-x4plus-anime.bin/.param   <- anime model (8.6 MB)
    realesr-animevideov3-x*.bin/.param   <- fast models (1.2 MB × 3)

Sources:

• Binary: xinntao/Real-ESRGAN-ncnn-vulkan v0.2.0
• Models: xinntao/Real-ESRGAN v0.2.5.0

---

Dependencies

pillow   -> only for reading image dimensions and mode (display info)

The actual upscaling is done by the NCNN binary compiled with Vulkan. No PyTorch, CUDA, ROCm, or heavy ML libraries required.

---

Why Real-ESRGAN NCNN Vulkan

Full research notes (alternatives evaluated, findings, what doesn't work on AMD Windows): see RESEARCH.md.

TL;DR: AMD GPU on Windows = Vulkan or nothing. NCNN Vulkan gives quality on par with Topaz / Aiarty without paying and without uploading anything to the cloud.

---

Credits

• Real-ESRGAN — Xintao Wang et al.
• Real-ESRGAN-ncnn-vulkan — MIT License

License

MIT — see LICENSE.