DLSS / FSR
AI / spatial upscaling for higher framerates
Also known as: DLSS · FSR · XeSS · Upscaling · Super resolution
Definition
DLSS (NVIDIA), FSR (AMD), and XeSS (Intel) are upscalers that render the frame internally at a lower resolution and reconstruct a higher-resolution image. DLSS uses an AI model and motion vectors; FSR 1 is spatial-only, FSR 2/3 is temporal like DLSS. For art teams, this means the engine resolves much finer detail than the raw render budget would suggest — but it also means LODs, normal maps, and shader detail need to hold up under aggressive temporal reconstruction.
In production
These upscalers render internally at a lower resolution — commonly 50-67% of target resolution in "Balanced" or "Quality" modes — and reconstruct a full-resolution image using temporal accumulation and motion vectors. DLSS uses a trained AI model running on NVIDIA's Tensor Cores, which is why it historically produced cleaner results than FSR 1, a purely spatial upscaler with no temporal data; FSR 2 and FSR 3 closed much of that gap by adopting the same temporal, motion-vector-driven approach, and both now ship as plugins directly inside Unreal Engine 5's post-process pipeline. For art teams, this means the engine resolves noticeably finer detail than the raw internal render resolution would suggest, but it also means LODs, decal transitions, and thin geometry need to hold up under aggressive temporal reconstruction, since upscalers can amplify ghosting or shimmer on assets not authored with temporal stability in mind. Frame generation (DLSS 3, FSR 3) adds a further wrinkle: entirely synthetic in-between frames mean UI elements and fast-moving VFX must be excluded via masks or they'll visibly smear.
