ALGOSYNTAX BLOG

VR Performance in Unreal Engine 5: How We Went From 20 to 68 FPS

November 18, 2026

Unreal Engine 5.8, PC VR through OpenXR. Updated November 2026. Written by the Algosyntax team.

Short answer: most of the speed came from turning off desktop features VR doesn’t need, not from clever code.

  • Learn how VR frame pacing works. A frame that misses its slot waits for the next one, so 72 Hz drops straight to 36 fps.
  • Profile with Unreal Insights on a packaged Development build. Don’t guess.
  • Turn off what costs extra in stereo: ray tracing, virtual shadow maps, distance fields, TSR, software VRS, SSAO, volumetric fog and clouds, real-time sky capture, cube-map shadows.
  • Count what’s on screen. Our biggest single win was showing 8 bars of falling notes instead of 64.
  • Put fixes in config or Blueprint defaults, never on one placed actor.

Result: from 18-24 fps to an average of 68 fps, with 86% of frames on time at 72 Hz.

This is the performance work behind VR Musician, our Quest 3 piano app: a MetaHuman pianist, a falling-notes highway and a live weather sky. All numbers here are from PC VR (VR Preview and a packaged Windows build on the Meta XR Simulator), not from a Quest. The method is what carries over.

1. Understand the frame budget first

At 72 Hz, the Quest 3 default, you have 13.9 ms per frame for the game thread and the same for the GPU. PC headsets often run 90 Hz, which leaves 11.1 ms.

The surprise for us: VR doesn’t slow down gradually. A frame that’s late waits for the next display slot, so the frame rate halves: 90 Hz becomes 45 fps, 72 Hz becomes 36 fps. Our “stuck at 43-45 fps” and “stuck at 36 fps” weren’t two different problems. They were both missed slots.

The second surprise: in stat unit, the Game time includes time the game thread spends waiting for the VR runtime (xrWaitFrame). In one trace, 9.3 ms of a 24.5 ms game thread was waiting. A high Game number doesn’t prove your gameplay code is slow. Check before you optimise Blueprints.

2. Check which scalability you’re testing

VR Preview uses the editor’s scalability settings, which default to Epic. At Epic we measured a 39-44 ms frame and about 4,500 draw calls. At Medium: 25 ms and 724 draw calls. Standalone Game from the editor has its own scalability setting, so changing the editor’s doesn’t change it.

3. Profile with Unreal Insights, and export it without the UI

stat game showed World Tick at 16 ms with only 2.8 ms named. Insights showed the rest. Record from a packaged Development build, because the editor adds its own costs:

				
					MyGame.exe -trace=default,counters,stats -statnamedevents -tracefile="D:/Traces/Playback.utrace"
				
			

You can turn a trace into CSV files without opening Insights, which is handy for comparing runs or handing them to an AI assistant. Put the export commands in a text file and run:

				
					UnrealInsights.exe -OpenTraceFile="D:/Traces/Playback.utrace" -NoUI -AutoQuit -ExecOnAnalysisCompleteCmd="@=D:/Traces/export.txt"
				
			

With a line like this in export.txt:

				
					TimingInsights.ExportTimerStatistics D:/Traces/game.csv -threads=GameThread -startTime=20 -endTime=120 -sortBy=TotalInclusiveTime -sortOrder=Descending
				
			

Skip the first 20 seconds or so: loading and shader compilation hide the steady-state costs. For the same reason, always measure a second, warmed-up run.

4. Turn off what VR doesn’t need

These went into Config/DefaultEngine.ini, so every level gets them. The mobile renderer on Quest has none of these features anyway, so the PC build also gets closer to what Quest shows.

SettingWhy
r.RayTracing=FalseNothing used it, but the engine still rebuilt ray-tracing geometry for the skinned MetaHuman every frame
r.Shadow.Virtual.Enable=0Virtual shadow maps render per eye
r.GenerateMeshDistanceFields=FalseOnly Lumen and distance-field AO used them (0.9 ms)
r.AntiAliasingMethod=2 (TAA)The default TSR cost about 2 ms in VR and softened the image
vr.InstancedStereo=True, vr.MobileMultiView=TrueBoth eyes in one pass on PC and Quest
r.VRS.EnableSoftware=0The engine turns software variable rate shading on for every Windows project; it cost 0.5 ms and saved nothing visible
r.AmbientOcclusionLevels=0SSAO, about 0.6 ms
r.Shadow.CSM.MaxCascades=2, r.Shadow.DistanceScale=0.5Sun shadows 1.7 → 0.4 ms for a seated scene
r.VolumetricFog=0, r.VolumetricCloud=0About 1.1 ms for fog, 2.8-3.4 ms for clouds. We draw a cheap cloud layer in the sky instead
r.SkyLight.RealTimeReflectionCapture=0Re-rendered the sky into the sky light every frame, about 1.5 ms
r.SupportPointLightWholeSceneShadows=FalseSee the next section

Lumen was already off: in VR it crashed with MetaHuman hair, and Quest doesn’t have it. Most of these settings are read at startup, so restart the editor before you measure. We once measured before restarting and still saw distance fields in the profile.

5. The invisible light that cost 5 ms

The trace showed a cube-map shadow drawn every frame: about 1.1 ms of GPU plus 4 ms of render work. It came from a faint, 1-candela fill light that nobody could even see next to the sun.

The project has static lighting turned off, so a light set to Stationary behaves like a Movable one and redraws its six-sided shadow every frame. We turned off Cast Shadows on that light, then turned off point and rect light shadows project-wide with the last setting in the table.

6. Count what’s on screen

Our note highway was 16 metres long with 25 cm per bar, so 64 bars of notes were visible: about 1,700 moving pieces. Shortening it to 2 metres (8 bars) at the same spacing kept the note speed and cut the pieces to about 200. Together with the cloud and sky changes, the GPU went from 12.9 to 5.3 ms.

The lesson applies far beyond music games: before optimising how each object draws, check how many you draw. A shorter view distance, a smaller pool or fewer spawned effects often beat any shader trick.

7. Fix the Blueprint, not the placed actor

Our second level ran at 30-36 fps while the first ran at 65-70. We had shortened the highway and fixed the shadowed light on the placed actors in the first level only. The second level, copied from our desktop app, still had the long highway, its own shadowed light, a volumetric cloud actor and real-time sky capture.

We moved the values to the Blueprint’s class defaults and the global settings above. Both levels then ran the same. If a performance fix lives on one instance, the next level won’t have it.

The result

StageResult
Start (Epic, everything on)18-24 fps
Short highway, no volumetric clouds, sky captured once, first config cutsAbout 64-66 fps in VR Preview
First packaged Development build, during song playbackAverage 53.4 fps, 64% of frames on time
After VRS, SSAO, volumetric fog, light shadow and cascade cutsAverage 68 fps, 86% of frames on time at 72 Hz, GPU 11.6 → 7.7 ms

What’s left is mostly things we chose to keep: MetaHuman hair and skin, lighting and anti-aliasing, each under about 1 ms. Our own code was cheap: the pianist’s animation and Control Rig took about 0.9 ms on worker threads, and the hands, panels and note highway under 0.4 ms each.

FAQ

Do these settings make PC VR look worse? A little: softer shadows, no volumetric clouds, no SSAO. In a headset, a steady frame rate matters much more than these features.

Is 60 fps enough? Not for a store release. Meta’s store review expects apps to hold the headset’s refresh rate, usually 72 Hz on Quest. Below that, frames halve and players feel it.

Does this tell me my Quest frame rate? No. The method is the same, but Quest uses the mobile renderer. Profile on the headset before release.

Related articles

← All blog posts