media-upscale

What problem does it solve?

AI super-resolution for images and video beyond ffmpeg's outdated built-in DNN models (SRCNN 2014 / ESPCN 2016 / EDSR 2017). Open-source + commercial-safe only: Real-ESRGAN (BSD-3-Clause, the current default for photos/video), Real-CUGAN (MIT, anime-optimized), SwinIR (Apache 2.0, transformer-based), HAT (Apache 2.0, 2023 SOTA), GFPGAN (Apache 2.0, face restoration), waifu2x-ncnn-vulkan (MIT, classic anime), Upscayl (AGPL-3.0, GUI wrapper around Real-ESRGAN), chaiNNer (GPL-3.0, node pipeline for arbitrary models). Realistic 2x / 4x / 8x SR for photos, film scans, low-res archives, anime, screenshots, game textures. Use when the user asks to upscale an image, super-resolution a video, enhance an old photo, restore a face, upscale anime, rescale game textures, do AI upscaling, or when ffmpeg's built-in SRCNN/ESPCN/EDSR isn't good enough.

Core Features & Use Cases

• Supports multiple open-source + commercial-safe models: Real-ESRGAN, Real-CUGAN, SwinIR, HAT, GFPGAN, waifu2x-ncnn-vulkan, Upscayl, chaiNNer.
• Provides an end-to-end pipeline for single images, videos (frame-by-frame), and batch processing via scripts/upscale.py.
• Includes guidance on model selection, installation hints, and license safety via references docs.

Quick Start

Invoke the driver script to upscale a single image or video by selecting a model and scale and providing input and output paths.