A research-grade upscaler that runs on your laptop GPU

What it does This is a standalone C implementation of Real-ESRGAN, the blind super-resolution model trained on synthetic data. It upscales and restores images — photos, compressed JPEGs, and anime — using Vulkan for GPU acceleration. No Python environment required; it runs as a native binary on Windows, Linux, and macOS.

The interesting bit The project sits on Tencent’s ncnn inference engine rather than PyTorch or ONNX Runtime, which is the unusual part. ncnn is built for mobile and edge deployment, so this gets you a desktop-grade upscaler with the portability expectations of a phone app. The authors also explicitly tuned for anime — a niche that general restoration tools often smear into watercolor mush.

Key highlights

• Ships with four pretrained models: general photo (realesrgan-x4plus), anime (realesrgan-x4plus-anime), video-optimized anime (realesr-animevideov3), and a lighter network (realesrnet-x4plus)
• Supports 2×, 3×, and 4× scaling; tile-based processing for memory-constrained GPUs
• Multi-GPU support via device IDs and per-GPU thread allocation
• TTA mode for the quality-obsessed (test-time augmentation, slower but sharper)
• Outputs PNG, JPG, or WebP; handles batch directory processing

Caveats

• The TODO list admits a bug where “some PCs will output black images” — cause unclear, driver updates suggested
• No face restoration yet; GFPGAN integration is planned but not implemented
• Model conversion guidance is missing; you’re stuck with the bundled checkpoints unless you already know ncnn

Verdict Worth a look if you want Real-ESRGAN without the Python dependency stack — think batch-processing workflows, embedded pipelines, or just avoiding conda. Skip it if you need face restoration, custom model training, or guaranteed stability on every GPU driver version.