RealScaler: a DirectML desktop wrapper around Real-ESRGAN for Windows

The gap RealScaler fills between Real-ESRGAN and a usable desktop app

Real-ESRGAN is a research codebase. Getting a result out of it means cloning a repository, choosing a weights file, writing an inference script, and then handling the parts nobody documents: what happens when a 4K frame does not fit in VRAM, how to keep audio when you process a video, how to resume after a crash on hour three of a batch. RealScaler exists to absorb those parts. The README describes it as a Windows app powered by RealESRGAN AI to enhance, upscale and de-noise photos and videos, and the feature list is essentially a catalogue of the operational problems that surround the model rather than the model itself. Target users are people who already know they want Real-ESRGAN output (anime frames, manga scans, compressed footage) but do not want to maintain a Python inference harness around it. The stated requirements are narrow: Windows 10 or 11, at least 8GB of RAM, and any DirectX 12 compatible GPU with 4GB or more of VRAM. There is no Linux or macOS path in the material.

How the DirectML engine, tiling and video pipeline fit together

The stack is entirely Python. PyTorch and ONNX appear in the dependency list, but the inference engine in the 3.X line is onnxruntime-directml, per the changelog entry "New AI engine powered by onnxruntime-directml". That choice is what makes the AMD, Intel and Nvidia claim possible: DirectML is a Windows abstraction over DirectX 12, so the app does not need CUDA or ROCm to reach the GPU. The 1.X notes mention an earlier switch to Pytorch-directml for the same reason. Models live as onnx files in an AI-onnx folder, and the README's manual setup step is to extract them there. For video, FFMPEG does frame extraction (the 4.X notes claim extraction is ten times faster than the previous approach), frames go through the model, and the result is re-encoded. Audio handling arrived in the 1.X line, and the 4.X line added hardware accelerated encoding through nvenc, amf or qsv. Two mechanisms deserve attention. Automatic image tilling splits large inputs so a frame that exceeds VRAM is processed in pieces, which is the standard mitigation but also the source of seam artifacts if tile overlap is wrong. STOP&RESUME lets a long video job be interrupted and continued, which matters because a multi-hour upscale on a consumer GPU will otherwise be lost to a driver reset. The GUI is CustomTkinter, and the app is packaged with PyInstaller; metadata is carried from source to output via exiftool.

Setting it up from source: the commands and folders that matter

There is no installer described in the README. The manual route is: download the project ZIP (Code > Download ZIP), extract it, put the AI model files into the AI-onnx folder, and put ffmpeg.exe into the Assets folder. The README points to a gofile.io link for the models and to gyan.dev for the FFMPEG release build (ffmpeg-release-essentials.7z). Open the project in VSCode, select RealScaler.py, then run pip install -r requirements.txt from the Terminal panel. The README then says to close and reopen VSCode to refresh dependencies and click the Play button. Note what is not in the repository: the onnx weights and the FFMPEG binary. Both are third-party downloads, and the model link is a file-hosting page rather than a release asset, so it can disappear independently of the code. The README also lists an itch.io page as the distribution channel, which is the route for users who do not want to assemble the runtime themselves. Supported inputs are listed as jpg, png, tif, bmp, webp and heic for images, and mp4, webm, mkv, flv, gif, avi, mov, mpg, qt and 3gp for video. The app saves user settings such as AI model and GPU selection between runs.

Where RealScaler is the wrong tool

The GPU requirement is the first hard boundary. The README asks for a DirectX 12 GPU with 4GB or more VRAM, so a machine with integrated graphics only, or a GPU below that memory floor, is out. Tiling exists to work around VRAM limits, but it does not remove them; it trades memory for time and introduces the possibility of visible tile seams on detailed content. Windows is the second boundary: DirectML is a Windows API and the entire dependency list (onnxruntime-directml, the nvenc/amf/qsv encoding paths) is Windows-shaped. Third, the model distribution model is fragile. Because the weights are not in the repository, a fresh clone is not a working install; if the gofile link rots, the manual setup path breaks and the only remaining route is the itch.io build. Fourth, this is a GUI application. There is no documented CLI or batch scripting interface in the material, so anyone wanting to wire upscaling into an automated pipeline would be fighting the tool rather than using it. Finally, the 2026.X roadmap shows TTA upscale mode still unchecked, so the highest-quality inference mode is not present in the current release line.

QualityScaler and the same author's other projects

The most direct alternative is QualityScaler, from the same author and linked in the README under "Other AI projects". It is also described as an image and video AI upscaler app, which means the difference is not the general category but the underlying model. RealScaler is explicitly built on Real-ESRGAN, and the README credits xinntao/Real-ESRGAN as the source. QualityScaler occupies the sibling slot in the same family, and FluidFrames covers video interpolation. That structure is worth understanding before choosing: if you want Real-ESRGAN output specifically, RealScaler is the wrapper for it. If you want a different upscaler architecture, or interpolation as a separate concern, the author has separate applications for those. Outside this family, the obvious alternative is running Real-ESRGAN directly from xinntao's repository, which gives you the full inference API, batch scripting and cross-platform support, at the cost of writing the frame extraction, audio muxing, resume logic and VRAM management yourself. RealScaler's value is precisely that it has already written those parts for Windows.

Release cadence, maintenance and the MIT licence

The release history shows three tagged versions: 4.1 in March 2025, 4.7 in October 2025, and 2026.3 in April 2026. That is roughly one significant release every six months, with the version numbering shifting from a plain 4.x scheme to a year-based 2026.x scheme. The changelog is written as a checklist of completed items per major version, and the 2026.X block already has two items checked (new AI models, interface scaling) with TTA upscale mode still open. For an adopter, that means the project is active but the upgrade path is not a package manager. You re-download the source or the itch.io build, and if a release changes the model set you may need to fetch new onnx files to match. The licence is MIT, which is permissive and permits commercial use and modification, but note that the licence covers the RealScaler code, not the Real-ESRGAN weights it downloads, which carry their own terms. That distinction is worth checking before shipping upscaled output commercially. This is not legal advice; read the upstream model licence yourself.

Editorial conclusion

Adopt RealScaler if you are on Windows 10 or 11 with a DirectX 12 GPU of at least 4GB VRAM and you want Real-ESRGAN output without touching a notebook or a CLI. Do not adopt it if you need Linux or macOS, if you have no GPU, or if you want a single binary you can install and forget, because the AI models and FFMPEG are separate downloads the README links to rather than bundled artifacts. Before committing, confirm that the model archive from the README's gofile link is still reachable and that the onnx files land in the AI-onnx folder, since nothing in the repository supplies them.

Official sources

1. Djdefrag/RealScaler on GitHub
2. Issues
3. License: MIT
4. README
5. Releases