Pro Edges reads your GPU, VRAM and CPU, then orchestrates 30+ open-source AI engines — Real-ESRGAN, SwinIR, RGT, NAFNet — to rebuild photos, anime and video at up to 16× resolution. Nothing leaves your machine.
Real output
A 150×100 degraded frame restored to 600×400 — captured from the exact pipeline you get after installing. Switch methods below the frame.
Capabilities
GAN-sharpness or PSNR-fidelity engines repair JPEG artifacts, noise and motion blur while scaling 2×–16×. Faces recover through dedicated restoration passes.
Frame-accurate batch processing through FFmpeg with audio preserved. Export H.264, ProRes, VP9 or PNG sequences at full upscaled resolution.
CUDA, DirectML and CoreML detected at launch — tile sizes adapt to free VRAM so nothing overflows.
Inference never leaves the device. The only network call fetches open-weight models you ask for.
Laptops run SPAN and ClearReality at realtime rates — 2 MB models, seconds per image.
The smart engine
At launch, Pro Edges builds a hardware fingerprint: GPU vendor, exact VRAM (driver registry — no 4 GB blind spot), free memory, CPU cores and RAM. Every recommendation is re-ranked against that profile in milliseconds.
$ proedges --scan gpu_vendor NVIDIA gpu_name GeForce RTX 4060 vram_total 8.0 GB (registry — true size) vram_free 7.6 GB backend DirectML cpu i7-12700F · 20 threads ram 31.8 GB score 63 / 100 tier HIGH # ranked for this machine, right now: recommended UltraSharp V2 q 8.5 alt_max SwinIR 4x GAN q 9.5 alt_fast SPAN 4x s 9.0
This exact scan powers the Recommended badges in the app — nothing hardcoded per device.
Model library
Each carries quality and speed ratings plus hardware-fit metadata. Your tier decides what appears in the app.
Best-in-class real-world detail recovery
DAT2 crispness, the all-round default
State-of-the-art denoising for brutal noise
Spatial pyramid attention, realtime-class speed
Free forever. Windows 10/11 · 64-bit · GPU optional (CPU fallback). No Python needed.
FAQ
No. All inference runs locally through ONNX Runtime or PyTorch on your own GPU or CPU. The application's only outbound connection fetches open-weight model files from public repositories when you request a new model.
On startup the engine profiles your machine: GPU vendor, total and free VRAM (read from the driver registry, so cards above 4 GB report correctly), backend availability (CUDA / DirectML / CoreML), CPU cores and system RAM. Each model carries quality, speed, architecture and VRAM metadata; recommendations combine curated order with a hardware-fit penalty — transformers deprioritized on CPU, for example.
Yes. DirectML works on any DirectX 12 GPU, covering modern Radeon and Intel Arc/Iris boards. Those devices are steered toward CNN-based models that run efficiently under DML.
Under 4 GB VRAM you get the light tier: SPAN 4x, ClearReality and AnimeVideo v3 produce strong results quickly with tiny downloads. Tiling adapts down to 128 px so memory limits are respected.
Yes — Denoise → Deblur → Detail → Face restore → Upscale runs as one pipeline, each stage using the best model your tier supports.
Yes. Audio is demuxed before frame processing and remuxed into the finished file.
The application ships exclusively through the signed installer channel on this page. The installer repository is public and auditable; the development repository stays closed.