Rendering Portals in Virtual Reality

Portals have many applications in the field of computer graphics. Recently, they have found use as a way of artificially increasing the available space in a virtual reality (VR) environment. In this paper, we will cover a technique for making the transition through a portal unnoticeable to the user. Additionally, we will measure the performance impact of rendering portals in a test scene and provide some insight into possible optimisations.

💡 Research Summary

The paper addresses a fundamental limitation of room‑scale virtual reality: the physical space available for a user to walk is far smaller than the virtual environments designers wish to present. Building on the concept of “impossible spaces,” where multiple virtual rooms are linked through portals that appear to occupy the same physical area, the authors propose a method to make the transition through such portals visually seamless. Traditional portals are rendered as flat quads; when a user’s head crosses the portal plane, the view abruptly switches, causing a noticeable flicker, especially in stereoscopic VR where each eye is rendered separately. To eliminate this artifact, the authors replace the flat portal with a three‑dimensional box whose six faces are textured with the view from the adjoining room. By exploiting back‑face culling, only the exterior faces are visible from outside the portal, while the interior faces become visible when an eye penetrates the box. This ensures that even if one eye enters the portal before the other, both eyes continue to see consistent imagery, producing an “unobtrusive” transition.

Performance implications are investigated through a controlled test scene containing three pairs of connected portals. The scene is rendered on an Nvidia GTX 1070 while varying the number of active portal pairs (0, 2, 4, 6). Results show a dramatic drop in frame rate—from 44.7 fps with no portals to 10.3 fps with six pairs—and a corresponding increase in GPU frame time from 0.3 ms to 78 ms. The authors attribute the majority of this overhead to the need for additional render passes: each portal pair requires rendering the scene twice per eye, inflating the total number of passes from two (no portals) to fourteen (six portals). Texture uploads for portal views also contribute significantly to the GPU load.

To mitigate these costs, two optimization strategies are presented. The first employs the stencil buffer to mask out only the screen region occupied by each portal. In the first pass, the portal geometry writes to the stencil; subsequent passes render the portal‑view texture only where the stencil permits, preventing unnecessary pixel work regardless of how many portals are visible. The second strategy is single‑pass instanced rendering, where both eyes are rendered in a single draw call using GPU instancing. This reduces geometry duplication and eliminates multiple texture uploads, but it also precludes independent per‑eye portal teleportation, representing a trade‑off between performance and visual fidelity.

The conclusion emphasizes that while the box‑portal plus back‑face culling technique successfully hides portal transitions, rendering multiple portals remains computationally expensive. The proposed stencil‑buffer masking and single‑pass instancing can substantially lower the cost, making it feasible to display several portals simultaneously in real‑time VR. Future work should explore hybrid approaches that combine per‑eye teleportation with efficient masking, dynamic resolution scaling for distant portals, and smarter visibility determination to further close the gap between physical and virtual space without sacrificing immersion.