Superseded by 0.14.0-beta.4 for the helper window resize: beta.3 resized the helper but the in-game panel stretched instead of re-laying out (the add-on re-handed its old panel texture). Everything else below stands.
You probably do not need to install this by hand. There is a one-command installer that does
everything the README's manual sections describe:
tools/Install-DLSS5Feeder.ps1.
Download it, right-click → Run with PowerShell, point it at your game folder. It fetches ReShade,
the shaders, the neural consumer and both NVIDIA runtimes, and switches (-Consumer,-LocalFiles,
-NoElevate,-Api) cover the awkward cases. SeeAUTOMATIC_INSTALLATION_AVAILABLE.txtin the
assets below.
IPC went from v7 to v8. 32-bit games must update BOTH dlss5-feed.addon32 and host64\dlss5-feed-host64.exe — a mismatched pair will refuse to talk.
The helper window is resizable now, and it means it
Asked for twice, and it still did not work. Both halves of "too narrow, and expanding it doesn't scale correctly" were true:
- The width/height sliders only wrote
host64\ReShade.ini, which the helper reads once, at startup. Moving them appeared to do nothing until you found "Restart the DLSS 5 host" further down the page. - The window has always had a resizable border, but nothing answered
WM_SIZE— so the swapchain kept its original size and Windows stretched it into whatever shape you dragged. The picture distorted; the UI never got more room.
There is now one real resize path, used by both the sliders and the border drag: the window, its swapchain, the banner and the panel texture cast into the game are all rebuilt, and ReShade's own docked tab column is re-fitted to match. The sliders apply immediately — no restart.
Measured end to end: 900x992 → 1384x720, banner and panel texture rebuilt, tab column re-fitted from 804 px to 1288 px, and --test still passes 300/300 afterwards.
- Default width 620 → 900. 620 left under 500 px for the neural consumer's panel once ReShade's own chrome was accounted for.
- Slider ranges now match what the helper accepts: 300–4000 wide, 0–8000 tall (0 = fill the work area).
- Still saved to
[DLSS5Host] WindowWidth/WindowHeight, so the choice survives a restart.
The in-game panel can move too. It was pinned to the game window's top-right corner with no control at all. "Panel corner" on the overlay (cast_anchor) puts it in any of the four corners.
D3D12CreateDevice failures now explain themselves
Three reports this week where the private D3D12 device failed inside the game while host64\dlss5-feed-host64.exe --test created one fine on the same machine, adapter and driver — each with a different code, all logged as bare hex.
- The codes are named.
0x887E0003isD3D12_ERROR_INVALID_REDIST,0x887A0004isDXGI_ERROR_UNSUPPORTED, and so on. Every constant checked against the Windows SDK headers. - The adapter is logged before the call, not only after a device exists (#47 turns on exactly this — a failed create used to say nothing about which adapter it tried).
- A game-local
D3D12\folder is reported. That is an Agility SDK redist path: if the game points Direct3D 12 at it and it is empty, every device created in that process fails, ours included. This is the likely cause of #61.Verify-DLSS5Feeder.ps1checks for it too. - The create is retried once with DRED disarmed. Arming DRED is the one thing the add-on does before creating a device that the helper does not — and the helper is the one that succeeds. If the retry works you keep a working session and lose only the crash breadcrumbs.
If you set DLSS5_FEED_D3D12_DEBUG=1: the D3D12 debug layer is known to make D3D12CreateDevice itself fail inside a game process. Clear it before reporting (#65).
A legible message instead of a shader-compiler error
On ReShade's DirectX 9 backend DLSS5_Feed.fx failed with error X3531: can't unroll loops marked with loop attribute, which tells you nothing. It means ReShade is running its D3D9 backend, so dgVoodoo2 is not in effect and the add-on could never have attached. The shader says that now (#56). A 64-bit D3D9 game does not need this project at all — renodx-dlss handles those alone.
The helper no longer races itself
If your helper takes a while to start — one reporter's needs about 11 seconds — the overlay button read "Start the DLSS 5 host" for that entire window, because the only thing it consulted could not tell a starting helper from a dead one. Pressing it killed the one on its way and spawned a rival for a named pipe only one of them can own. The button is now disabled and says "The DLSS 5 host is starting…", and a restart request in that state is refused (#58).
Vulkan
- Detroit: Become Human's 64-bit Vulkan flicker is fixed — contributed by @danzig666 as PR #66, with a deterministic two-queue GPU regression test. The capture could run before the game's render work finished; the present waits are brought forward and re-signalled. (#13, #8)
- A follow-up fix to that same code: its present hook is installed per Vulkan device but was only removed when the add-on unloaded. A game that destroys its Vulkan device and creates another without destroying the instance left the old hook behind and never hooked the new one.
Not fixed here
- #62 (X4 Foundations Vulkan crash) — the dumps are being read. That report is on 0.13.1-beta.1, which predates PR #66, so a beta.3 re-run is worth doing first.
- #47 — still open. The driver, architecture and transport theories have all now been eliminated by counter-example; the logging above is to give the thread better evidence, not a fix.