The difference between AR and VR comes down to what happens to the real world. Augmented reality (AR) keeps your physical surroundings in view and attaches digital information to objects and places in them. Virtual reality (VR) replaces those surroundings with a world the computer synthesises, built so that you feel you are really present inside it. So an AR system needs more knowledge of the room it is in, while a VR system offers high immersion with little or no world knowledge.
AR, VR and mixed reality on the reality–virtuality continuum
Since 1994, AR and VR research has been framed by Milgram and Kishino’s reality–virtuality continuum. It places displays on a continuum running from a fully real environment to a fully virtual one, and treats any display that presents real and virtual objects together as mixed reality. Their working definition of augmented reality covers any case where an otherwise real environment is “augmented” with computer-graphic objects. Mixed reality, the middle of the scale, holds both augmented reality, where the virtual adds to the real, and augmented virtuality, where real elements are brought into a mostly virtual world. Augmented virtuality is the mirror image of AR: it embeds live real-world content in a virtual space rather than placing virtual content in a real one.
A later paper revising the continuum pushes the model further. It argues that the “virtual reality” endpoint can never be reached, because the headset is itself a real object sitting in a real room, so every technology-mediated experience is really mixed reality. Mixed reality is still the least precisely defined term in the family and the one procurement teams confuse most often. If the term you are trying to pin down is MR rather than AR or VR, our mixed reality examples show what holds up in real use.
World knowledge versus immersion: the split that drives the engineering
Milgram and Kishino separated display systems along three dimensions: extent of world knowledge, reproduction fidelity and extent of presence metaphor. The 2021 Frontiers paper folds the last two into one dimension, immersion. On those axes, conventional VR systems have high immersion and little or no world knowledge, while AR systems have low or medium immersion and much more world knowledge. Put simply, VR renders its world in fine detail but takes no account of the real environment.
The same paper draws a second distinction. In VR, coherence is internal: do virtual objects behave predictably with one another and with the user? In AR, coherence is external: a virtual object meant to sit on a table is broken if it floats in the air instead. When you move, the digital information should stay with the object or place it enhances, as if tied to it by invisible threads.
Registration, not rendering, separates AR from a plain graphical overlay, and it is where most production AR systems succeed or fail. For the full engineering picture of what augmented reality is and how its registration pipeline is built, see our dedicated guide. There is also a shorter overview of what augmented reality is and where it is applied.
How do AR and VR differ in hardware, from phones and tablets to headsets?
AR runs on phones as well as headsets. Mobile AR overlays text, images, audio and 3D content on the phone’s camera view of the physical environment. Current phones and headsets carry high-resolution cameras, inertial measurement units, GPS and, increasingly, LiDAR, all of which feed location and orientation estimates to AR frameworks. Google ARCore and Apple ARKit expose per-pixel depth maps, which AR apps use for occlusion and depth cues. VR is delivered mainly through head-mounted displays (HMDs).
A survey of workplace XR research found HMDs to be the preferred output device for both VR and AR. In VR the reason is immersion compared with CAVE rooms or flat screens, and in AR it is hands-free operation, which a tablet cannot offer.
Optical versus video see-through
Inside AR headsets there is a further split, between optical and video see-through. An optical see-through HMD shows the real world through half-transparent mirrors that also reflect the rendered image into the eye. A video see-through HMD captures the world with head-mounted cameras and combines it with graphics electronically. Optical designs leave the real view nearly intact, while video designs block the real view in exchange for merging the two more convincingly.
These trade-offs carry real consequences:
- Failure mode. If an optical see-through system fails, the user keeps their view of the real world. If a video see-through system fails, the user can lose both the real and the virtual view.
- Occlusion. Given a depth map, video see-through can block either scene pixel by pixel, whereas optical beam splitters have no way of selectively blocking the real environment behind a virtual object.
- Field of view. In Rolland and Fuchs’s 2000 review, optical see-through HMDs typically gave a 20 to 60 degree overlay field of view, and video see-through optics 20 to 90 degrees. For comparison, a VR reference table from a 2020 arXiv paper lists 120 degrees horizontal for a stereo VR headset.
- Resolution. Road to VR’s report on the HoloLens 2 announcement gives 47 pixels per degree, the level needed to read 8-point text, against about 16 for the Vive Pro VR headset of the time. Field of view and pixel density pull against each other, since stretching one dilutes the other.
The line is now a compositing mode as much as a device
The AR/VR distinction also appears as three environment blend modes. Opaque mode shows rendered layers with no view of the physical world and is the typical VR mode. Additive mode adds light over the real world, so black pixels turn transparent, which is typical of see-through AR headsets. Alpha-blend mode mixes rendered pixels with a camera view of the world, which is typical of phones and passthrough headsets. Opaque mode can also run on passthrough devices. That is why one headset can deliver both experiences: Meta’s documentation describes stereo cameras feeding the real world to the display while the app renders on top, with Quest 3 adding depth estimation so physical objects can occlude virtual ones.
What AR and VR demand from the GPU
One caveat before the figures. None of the sources measures AR and VR GPU load on the same workload side by side.
VR: two views, high refresh, a hard deadline
Traditionally a VR application draws its geometry twice, once for each eye, which NVIDIA’s VRWorks documentation describes as essentially doubling the processing of a non-VR application. Techniques such as NVIDIA’s Single Pass Stereo on Pascal-generation GPUs draw the geometry once and project it to both eyes, which frees budget for roughly double the geometric detail.
The rendering deadline is the other constraint. Typical XR hardware requirements, as reported in industry specs, call for sustained 72 to 120 fps stereo at under 20 ms motion-to-photon. That leaves none of the slack a flat-screen game can lean on. At 90 Hz the frame lasts 1000 / 90 ≈ 11.1 ms, and a 2025 arXiv study on standalone headsets cites earlier research putting the working budget at about 11 ms to render both eye buffers. A comparison table in the 2020 arXiv paper on multi-GPU VR rendering lists a 5 to 10 ms frame latency target for stereo VR, against 16 to 33 ms for PC gaming. The same paper notes that a missed deadline drops a frame, and that reprojection techniques such as Asynchronous Time Warp fill the gap without solving the underlying problem.
Measured latency changes over the course of a movement. One latency study tested the HTC Vive, Oculus Rift CV1, Oculus Rift S and Valve Index. It measured mean latencies of 21 to 42 ms at the start of a sudden movement, which fell to 2 to 13 ms once motion prediction took over. A measurement at a single point in a movement does not capture headset latency, because of motion prediction. Among vendor figures, Apple specifies 12 ms photon-to-photon latency and refresh rates of 90 to 120 Hz for Vision Pro, which is Apple’s claim rather than an independent test.
AR: the cost sits in tracking, stability and the camera feed
In AR, latency shows up as registration error. Rolland and Fuchs identify lag between measuring head pose and presenting the matching image as the largest source of registration error, at 60 to 180 ms in typical systems of their day. Optical see-through cannot delay the real scene to match the graphics, so those systems need low latency and predictive tracking. Microsoft’s HoloLens guidance makes a stable frame rate the first pillar of hologram stability: developers must deliver a new image every 16 milliseconds (60 FPS), with hardware-assisted reprojection correcting for head motion.
Tracking is the other cost. The workplace survey reports that tracking in augmented environments tends to be less stable than in VR, and harder to implement because it depends on fiducial markers or distinctive features in the surroundings. Mobile AR uses either marker-based or markerless tracking. Markerless approaches such as SLAM build a 3D map of the environment while tracking the device inside it. Heavy perception work does not always fit on the device: one mobile AR survey notes that object detection and recognition typically run on servers, because client devices lack the compute and battery to sustain them.
Passthrough mixed reality on a VR headset adds sensor processing to the rendering load. The 2025 Quest 3 study reports 17% lower GPU and 14% lower CPU performance in mixed reality mode than in VR-only mode, because of sensor processing overhead. It also estimates that MR applications may use up to 40% more GPU than their VR-only equivalents. These figures come from a simulation-based feasibility analysis. For how unified virtual memory affects XR rendering, see our piece on unified virtual memory and XR rendering budgets.
AR vs VR comparison at a glance
| Dimension | Augmented reality (AR) | Virtual reality (VR) |
|---|---|---|
| Relation to the real world | Adds virtual content to a real environment | Replaces it with a synthesised world |
| Taxonomy position | Low or medium immersion, higher world knowledge | High immersion, little or no world knowledge |
| What “coherent” means | Virtual objects behave correctly with real ones | Virtual objects behave correctly with each other |
| Typical devices | Phones and optical or video see-through HMDs | Head-mounted displays |
| OpenXR blend mode | Additive or alpha-blend | Opaque |
| Engineering pressure point | Tracking stability and lag-driven registration error | Doubled stereo geometry and a hard frame deadline |
AR or VR for a project: the trade-offs that decide it
A retail try-on, a surgical training simulator, an industrial remote-assistance overlay and a virtual museum reconstruction are not the same project at different scales. Each picks a different paradigm because the coupling to the environment, the session length and the content authoring differ.
Choose AR when the task happens in the real world and the user needs their hands, their tools and their surroundings. The workplace survey highlights AR and MR as an asset in low-visibility environments, for people in protective suits, and wherever conventional input devices are impractical. The costs are tracking that is harder to stabilise, a narrower overlay on see-through optics, and registration error from tracking lag. Choose VR when the environment itself is the content: somewhere dangerous, expensive or impossible to stage, where shutting out the room is the point. The costs are doubled geometry work, an unforgiving frame deadline and a user who cannot see their surroundings.
Session length narrows the choice further. Apple lists up to 2.5 hours of general use on Vision Pro’s battery and a device weight of 750 to 800 grams. The Quest 3 study estimates its native 720p30 MR compositing pipeline could run only 5 to 10 minutes before thermal throttling. Evaluation criteria differ too: for an automotive AR head-up display, being useful and safe matters more than being highly world-aware. Once the difference is clear and the real question is which paradigm fits, see our guide to choosing AR, VR or MR for a use case. For VR specifically, see our collection of VR use cases and examples.
Does VR teach better than AR? What the training evidence shows
One study compared learning in a VR condition with a non-VR condition. Students reported feeling more present (d = 1.30) but learned less (d = 0.80), and showed higher cognitive load on an EEG measure (d = 0.59). Presence is an experience metric, not a learning outcome, and one study of one lab simulation cannot settle the question for every course.
AR training works differently. It overlays guidance on real objects. In one neurosurgical training tool, a tablet shows AR guidance for placing a ventricular drain on a mock-up head, with a haptic stylus giving feedback. In another, trainees in AR headsets see guidance superimposed on a patient manikin while using real medical instruments. In military training and simulation, virtual scenarios replace live-fire drills. We would decide on the skill being trained: procedural work with real tools leans towards AR, while rehearsing rare or hazardous situations leans towards VR.
FAQ
What is AR vs VR vs MR vs XR?
Extended reality (XR) is the umbrella term for AR, MR and VR, all placed on Milgram’s reality–virtuality continuum. Mixed reality covers everything between the fully real and fully virtual ends, including AR and augmented virtuality.
What is the difference between AI and virtual reality?
Virtual reality is a display and interaction medium that immerses the user in a synthesised world, while AI is software that perceives, predicts or decides from data. In AR, object detection and recognition routines locate what the camera sees so virtual annotations can be placed and tracked.
