Repository navigation
There's no installer to download yet: build it from source with BUILDING.md (git checkout windows-v0.2.1). Full history in CHANGELOG.md.
Released for: Windows (windows-v0.2.1). This is the first Windows release since 0.1.5, so it also brings
everything in 0.2.0 to Windows. Mac: not released, the latest Mac release is still 0.2.0. Nothing in this version
changes what happens on a Mac.
Upscaling on AMD and Intel graphics
Why: until now the Windows version needed an NVIDIA card. Many PCs have an AMD or Intel card instead.
What changed
- On a PC without an NVIDIA card, the upscaler and the flicker smoothing (RAFT) run on any DirectX 12 graphics card
through ONNX Runtime with DirectML (ort_models.py). The first upscale exports both models to ONNX once
(saved inmodels\onnx). The rest of the upscale step stays on the processor: it's light next to the models. - torch-directml was tried first and isn't used: it crashes on the upscaler's PReLU layers and inside RAFT.
- Measured on a 180-frame tape clip, running through DirectML on the test PC's card: 4.0 frames a second, against
6.7 with PyTorch on CUDA and 0.4 on the processor alone. The picture matches the CUDA one (66.8 dB PSNR). - Without
onnxruntime-directmlinstalled, or if it fails to load, the upscale step falls back to PyTorch on the
processor exactly as before (slow, but it works). - The setup checks and Settings → Performance list the graphics card whatever its make. An AMD or Intel card
passes the check when DirectML is ready, and the check says what to install when it isn't. - BUILDING.md has the setup for AMD and Intel cards: the processor build of
PyTorch,onnxruntime-directmlandonnx.THIRD_PARTY.mdlists both.
Not changed yet: on AMD and Intel cards the face models still run on the processor, which is very slow (turn
Face restoration off for long videos), and the heat pause only reads NVIDIA cards.
0.2.0 on Windows
The heat pause and the computer details in Settings (both from 0.2.0) were checked on a Windows PC for this release:
Settings shows the processor, graphics card, memory, what the AI models run on, free space and the card's
temperature read through nvidia-smi. A full tape-mode restoration of a short clip with faces and sound on finished
with every check passing. The card never got hot enough to trigger a heat pause during testing; that part was
tested on a Mac in 0.2.0.