20 Releases 35 Tags

RSS Feed

• v0.23.0 7966bf2dd8

  Compare

  DarkRoom 0.23.0 Stable

  dtourolle released this 2026-10-05 02:28:44 +00:00 | 9 commits to master since this release

  The neural models — face indexing, AI denoise, scene grading, masks, the panorama fill — now run on an Intel GPU, try any other GPU through WebGPU, and use every processor core on Windows and in the Flatpak, where until now they ran on one.

  Intel graphics run the models. On an Intel integrated or Arc GPU the models now run through OpenVINO. Measured on a Raptor Lake laptop's Iris Xe, model by model against the same laptop's processor: 2.5× faster face detection with the thorough detector, 3.4× faster scene grading, 4× faster AI denoise with Fast and nearly 6× faster panorama border filling. The app itself chose the Iris Xe there at first launch, in its own log. The first launch prepares each model for the GPU in the background, a few seconds each, and the processor serves meanwhile. Linux needs Intel's OpenCL driver (intel-compute-runtime on Arch). Windows has that driver with the graphics driver, and the installer carries the runtime, but the Windows path has not yet run on Windows hardware.

  Any other GPU is tried through WebGPU. A graphics card no vendor-specific path covers — an AMD card on Windows or without ROCm, a phone's Mali GPU — is tried through WebGPU, on Vulkan or Direct3D 12. The app times it against the processor at first launch and keeps whichever is faster. It is unmeasured on that hardware so far; where it loses, nothing changes.

  Every core on Windows and in the Flatpak. The Windows installer and the Flatpak now carry ONNX Runtime, so even without a usable GPU the models run on every core — 8 to 10 times faster than the single core they had before. The Arch package carries it too and no longer needs onnxruntime-cpu. Each desktop package grows by about 110 MB for the two runtimes it ships.

  The right runtime when several are installed. A machine can now hold the bundled Intel and WebGPU runtimes beside a CUDA runtime fetched for an NVIDIA card or the distribution's ROCm build for an AMD one; the app loads the one that fits the graphics card, so an NVIDIA or AMD card keeps its own faster path.

  Phones without a Qualcomm chip. The APK carries a second, generic runtime (32 MB) for phones the Hexagon path cannot serve, and tries their GPU through WebGPU — untested on such a phone so far. A Qualcomm device never loads it, and a phone that does not say who made its chip is treated as Qualcomm's, as before.

  A steadier choice of accelerator. The first-launch timing now warms the GPU up before measuring it — a cold integrated GPU lost to the processor in two tries of three — and Settings names the accelerator that lost to the processor, not one the runtime was never built with.

  Install over 0.22.1 as usual; the first launch re-measures the accelerators once.

  The APK, the desktop binary, the Windows installer and SHA256SUMS are built by CI from this tag's commit.

  Downloads

  • DarkRoom-0.23.0-x86_64-setup.exe
    262 MiB

    2026-10-05 13:25:08 +00:00
  • SHA256SUMS
    298 B

    2026-10-05 13:25:28 +00:00
  • darkroom-0.23.0-arm64-v8a.apk
    581 MiB

    2026-10-05 13:25:20 +00:00
  • darkroom-desktop-0.23.0-x86_64-linux
    148 MiB

    2026-10-05 13:25:28 +00:00
  • Source Code (ZIP)
  • Source Code (TAR.GZ)