Offline rendering can simply throw more computation at a problem. Real-time can't. That constraint changes everything about how you approach lighting.
This is a take that tends to land awkwardly when I mention it to someone else: real-time lighting is harder than offline rendering.
Offline rendering is what the film industry uses. It produces photorealistic results that real-time can't quite match — though the gap is getting closer, with hardware ray tracing and more material standardization becoming standard in games. Offline path tracers simulate millions of rays per pixel across multiple bounces, catching every caustic, every subtle interreflection, every gradient in a shadow, and penumbras across assets of wildly different scales. How could anything harder than that exist?
The answer is that offline rendering solves a physics problem — over hours, if needed. Real-time rendering solves the same physics problem inside a budget of approximately 16 milliseconds per frame at 60fps. Those are not the same challenge.
Offline Gets to Be Honest
In an offline renderer, the path tracer does something elegant: it simulates light transport as it actually occurs in real life. Rays bounce, scatter, and accumulate energy; the solution is noisy early and accurate late, and you simply wait for it to converge. Add more samples and the result improves. Add more time and it improves further. The algorithm doesn't know or care how long it runs.
This is a fundamentally different problem from real-time. The physics doesn't get simplified or approximated — it gets sampled more or less densely. Errors are statistical noise that averages out, not structural approximations that introduce systematic bias. Offline rendering can portray materials accurately: ground-truth subsurface scattering, physically correct glass refraction and diffraction, volumetric light transport — all tractable given enough time.
Real-time rendering has no such luxury. The frame must finish in 16ms at 60fps. Every technique used in that budget is an approximation of something the path tracer would compute correctly given enough time.
Real-Time Is an Exercise in Controlled Deception
The techniques that make real-time lighting work are, in a real sense, tricks. They are tricks built on deep understanding of the physics, but they are tricks.
IBL pre-filters incoming radiance from all directions into an environment cubemap and pre-computes the BRDF response into a lookup table, then reconstructs the full lighting integral with just two texture samples at runtime. This is the split-sum approximation, introduced by Brian Karis at Epic. Path tracing would evaluate that integral directly per frame. The precomputed version is fast and close — but it's static, it doesn't respond to scene changes, and the split-sum approximation introduces visible error at extreme roughness values.
Screen-space ambient occlusion samples a sphere or hemisphere of depth buffer values — typically at half resolution — and uses them to estimate occlusion. It's fast and visually convincing, but it misses geometry outside the screen, produces halos around silhouettes, and has no notion of which parts of the environment the occluded surface actually sees. Bent normals help, but a path tracer would simply trace the rays.
Shadow maps project geometry from the light's perspective and check depth values. They work, and they're everywhere. They also alias at range, have resolution budgets, can't easily handle area lights, and produce contact shadows only where the resolution allows it. Path-traced shadows are free once the rays are in flight.
Reflection captures bake a static cubemap at a point in the scene and use it for specular reflections. It is, in effect, a photograph taken during level build that gets composited onto every surface in the vicinity. Screen-space reflections layer on top and add dynamism, but they miss geometry off-screen and fall back to the static capture at the edges. A path tracer would just trace the reflection ray to wherever it terminates.
The Skill Requirement Is Different
Offline lighting lets you be relatively naive about the physics and still get correct results. Point a camera, place some lights, let the renderer do the integration. The errors are convergence errors, not systemic ones.
Real-time requires you to understand the approximations well enough to stay inside the range where they hold up. You need to know that IBL breaks down when the environment is high contrast and the surface is a rough conductor. You need to know where screen-space reflections will fail and what the fallback looks like. You need to understand why a light probe placed in the wrong position introduces color contamination across an entire room.
Getting real-time lighting to look correct isn't about understanding the physics. It's about understanding a stack of approximations layered on top of the physics, knowing where each one breaks, and composing them in a way that hides the seams.
That's a harder skill to develop than understanding the physics alone. And it's one that doesn't get enough credit, in my opinion.
© 2026 Stefan Groenewoud - All views are my own, not those of my employer.Â