github lonewolf0622/HelixSR v3.4.0
HelixSR 3.4.0

48 minutes ago

HelixSR 3.4.0. Download HelixSR Desktop below (Windows .exe, Linux / Steam Deck .AppImage or .deb) and let it set
HelixSR up, or use HelixSR-3.4.0.zip with its setup scripts. HelixSR Scaler 2.3.0 is included (its own download:
HelixSR Scaler releases), now with HelixSR custom frame generation.

New in 3.4.0

  • HelixSR custom frame generation (new). 2x up to 6x (one to five generated frames per real one). On a BC-250 at 1080p
    it takes about 2.6 ms for the one generated frame of 2x and 5.7 ms for all three of 4x (QuickFlow = fastest, the default:
    the in-between frames are made at half size and sharpened with the real frames' detail; off: full size, 4.3 / 9.6 ms).
    It ships in HelixSR Scaler's folder (OptiScaler/custom/HelixSR-FG) and works in games with DLSS Frame Generation: set
    FGInput=nvngxfg in OptiScaler.ini (HelixSR Scaler points FGNvngxReplacement at it already). Settings in
    OptiScaler/custom/HelixSR-FG/dlssg/helixsr.ini. Tested in Final Fantasy VII Rebirth (4x) on a BC-250 under Linux.
  • HLX DLSS DeepReuse ([Upscaling] DeepReuse = exact | 1:2 | 1:4 in HelixSR/helixsr.ini): the deeper layers run every
    second or fourth frame. BC-250 at 1080p: 1.02 ms (exact, default) -> 0.84 ms (1:2) -> 0.79 ms (1:4); slightly softer in fast
    motion. Off at Ultra Performance and in the Vulkan build. Tested on a BC-250 only.
  • NVIDIA GTX 16: HLX DLSS runs its 16-bit path by default with shaders made for Turing's 8-bit math, and HLX FSR 4 has a new
    version for GTX 16 (tier 10). GTX 10 and other cards are unchanged; the speed test times both paths on every card.
  • Preset K on Windows is faster: on AMD's Windows driver its outer layers keep their weights in constant buffers and its
    deep layers run as new DirectX 12 shaders. RX 6700 at 1080p: 16-bit 5.1 -> 3.0 ms, 8-bit 3.2 -> 2.5 ms. Linux keeps its
    Vulkan kernels (faster there); the BC-250's picture and speed are unchanged.
  • HLX FSR 4 Full for every RDNA card in one file (no AMD FSR file next to it). RDNA 2 (RX 6000), RDNA 3 (RX 7000) and now the
    RX 5300 / 5500 run the INT8 network with HLX FSR 4's own int8 kernels; the RX 5600 / 5700 get faster first and last passes on
    Windows; on a BC-250 it is 1-3% faster. New setting Int8Passes (which passes use the own kernels). Tested on a BC-250 and an
    RX 6700; RX 7000 and RX 5000 cards are untested.
  • Speed test: also times HLX FSR 4 Full's INT8 passes with its own kernels and with AMD's shaders, and Preset K's Windows
    kernels on and off, so each card can get the faster one.

Known issues in 3.4.0

  • Frame generation: HelixSR Scaler's menu (Insert) is interpolated too while it is open, and fast camera turns can leave a
    faint halo around characters.
  • A game can freeze at startup with Preset K when the Steam overlay is on. Turn the overlay off for that game (Properties, General) if it does.
  • Preset K with DeepReuse 1:2 or 1:4: one GPU hang on a BC-250 at 1440p with 1:2, cause not found. Exact, the default, is the safe choice.
  • HLX FSR 4 Full with 1:2 or 1:4: some ghosting in motion, more at 1:4. Exact, the default, has none.
  • Preset K can sparkle on foliage at any DeepReuse. HLX FSR 4 Full and Model E are steadier there.

Tested on an AMD BC-250 under Linux, plus Preset K and HLX FSR 4 Full on an RX 6700 under Windows. RX 7000, RX 5000, the Steam
Deck, NVIDIA GTX and Intel are not tested by us: reports with helixsr.log / OptiScaler.log, or HelixSR Desktop's speed test
results, are very welcome.

Don't miss a new HelixSR release

NewReleases is sending notifications on new releases.