EutropicAI's Final2x Unlocks Custom AI Upscaling Models on Your Own GPU
Final2x v4.0 ships a new PyTorch backend with custom model support and Nvidia 50-series GPU acceleration, all for free
Takeaways
- v4.0 ships a new backend: Final2x now uses the cccv PyTorch inference library, replacing the old ncnn/Vulkan core.
- Custom model support: Drop any PyTorch checkpoint into a YAML config and it appears in the GUI dropdown instantly.
- Nvidia 50-series GPUs: v3.0 added support for the latest Nvidia hardware; v4.0 builds on top of that.
- Bundled models: RealESRGAN, RealCUGAN, Waifu2x, and SRMD cover general photos, anime art, and degraded images.
- Runs fully locally: No cloud, no API key, no usage limits -- free and BSD 3-Clause licensed with 7.2k GitHub stars.
- Cross-platform: Available as a GUI binary for Windows, macOS, and Linux; also usable headlessly via
pip install cccv.
Final2x is a free, open-source desktop app that runs AI super-resolution models locally on your machine to upscale images to arbitrary sizes. With over 7,200 GitHub stars, it has quietly become one of the most popular local upscaling tools around. The project just shipped two major releases in quick succession: v3.0.0 added support for Nvidia 50-series GPUs, and v4.0.0 introduced a brand-new inference backend that unlocks custom model loading for the first time.
What changed in v4.0
Final2x v4.0.0 is now available, and it uses the new cccv backend, supporting custom models and more. The cccv library is a PyTorch-based inference engine built by the same team. It handles model loading, automatic padding, and GPU dispatch, and it supports three task types: single-image super-resolution (SISR), video super-resolution (VSR), and video frame interpolation (VFI).
The shift from the old Final2x-core backend (which used ncnn and Vulkan) to cccv (which uses PyTorch directly) is the real architectural change here. It makes adding new model architectures much easier, and it's what enables the custom model feature.
The models under the hood
Final2x allows image super-resolution to arbitrary sizes using multiple models, and currently supports RealCUGAN, RealESRGAN, and Waifu2x. Each targets a different use case: