• DarkRoom 0.24.0
    Benchmarks / CPU and I/O (per commit) (push) Successful in 8m56s
    Benchmarks / Frame budget (on demand) (push) Skipped
    Traceability / Requirement traces (push) Successful in 1m27s
    Build and test / Desktop (Linux) (push) Successful in 1h33m5s
    Build and test / Layer separation (push) Successful in 32s
    🐳 Android image / Build and push (push) Successful in 3s
    Build and test / android-image (push) Successful in 4s
    🐳 Windows image / Build and push (push) Successful in 1s
    Build and test / windows-image (push) Successful in 2s
    Build and test / Android (aarch64) (push) Successful in 45m54s
    Build and test / Windows (x86_64, cross) (push) Successful in 55m18s
    Build and test / Publish the release (push) Successful in 1m33s
    Stable

    dtourolle released this 2026-10-07 11:27:59 +00:00 | 0 commits to master since this release

    AI denoise with Best now takes about a second for a 20-megapixel photograph on a laptop's graphics card instead of three, with a new Best network and the whole frame denoised at once.

    A new Best. Best is now a single network, taught by the pair of networks that was Best until now, with a quarter of its training drawn from the sharp edges of real photographs. On real photographs it keeps edges as sharp as the old Best, a little sharper at high ISO, and its detail measures within 0.04–0.06 dB of it; the one place it falls short is smooth areas at the very highest ISO (25600), by 0.27 dB. It does a third of the old Best's work.

    Medium is gone. The methods are Bilinear, Fast and Best. A photograph last edited with Medium or the old Best opens with the new Best; nothing needs converting.

    The whole frame at once. On an NVIDIA graphics card, AI denoise now runs over the whole photograph in one or two large pieces instead of dozens of small tiles whose overlapping borders were computed and thrown away. The result differs from the tiled one by less than one step of an 8-bit image. Measured on a laptop's RTX 3050 with a 20-megapixel Canon 6D raw, network time:

    • old Best in tiles (0.23): 2.60 s
    • new Best in tiles: 0.95 s
    • new Best, whole frame: 0.51–0.54 s — about a second with reading the file

    First launch takes longer on NVIDIA. The graphics card prepares the whole-frame network once after installing, in the background — about 12 minutes on the RTX 3050 — while photographs are denoised the old way meanwhile. A card with less than 6 GB of memory cuts the photograph into more pieces, and a photograph that does not fit is planned again in smaller ones rather than failing.

    Other graphics cards and the tablet keep the tiled way, with the new network: Intel, AMD and Apple graphics, the processor alone, and the tablet's NPU, whose form of the new network measured lossless in simulation and is not yet confirmed on the tablet itself.

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

    Downloads