DLSS and FSR on cloud gaming work better than the marketing suggests

DLSS, FSR, and XeSS render the game at a lower internal resolution (1080p, 1440p) and use temporal AI upscaling to reconstruct a higher-resolution output (1440p, 4K). The benefit is roughly 1.5-2x the frame rate at the same visual quality, or roughly equivalent quality at a higher frame rate.

These techniques were designed for local rendering, where the upscaling pass runs on the same GPU that's doing the game render and the upscaled output is displayed locally on the user's monitor.

On cloud, the upscaling runs on the cloud GPU. The upscaled output is then encoded and streamed to the user. The video encoder takes the upscaled frame, compresses it, and sends it down the pipe at the cloud service's bitrate budget.

The relevant question: does the video encoder preserve the visual benefit of the AI upscale, or does it compress away the detail that DLSS/FSR added?

We compared cloud-streamed sessions with and without DLSS enabled across Cyberpunk 2077, Hogwarts Legacy, and Alan Wake 2 on GeForce Now Ultimate. The cloud GPU is an RTX 4080, encoder is AV1, output is 4K at 75 Mbps.

DLSS Quality vs native 4K: the streamed video is visually indistinguishable in motion. Both show similar levels of compression artifact in fast motion. The DLSS version delivers higher frame rate (consistent 90+ fps vs 50-60 fps native) and the streamed motion smoothness is meaningfully better as a result.

DLSS Performance (more aggressive upscaling) vs native: the streamed result is also indistinguishable, with even higher frame rates. The upscaling artifacts that would be visible on a local monitor (slight ghosting in foliage, occasional shimmer on fine detail) are compressed away by the video encoder before they reach the client.

Cloud gaming actually masks DLSS artifacts. The same upscaling that's visibly imperfect on a local monitor is invisible in a cloud stream because the video encoder's compression artifacts are larger than the upscaling's reconstruction errors.

Translation: DLSS Performance on cloud looks the same as DLSS Quality on cloud. On local hardware, DLSS Quality is the right setting for visual fidelity. On cloud, you can run DLSS Performance, get higher frame rates, and the streamed quality is unchanged.

This is the cleanest 'cloud is actually fine' finding for any specific feature category. AI upscaling on cloud is genuinely good and the marketing has not caught up to this fact.

Lower-bitrate cloud streams (Game Pass Cloud at 20 Mbps, Boosteroid at 15 Mbps) compress more aggressively. At those bitrates the encoder takes more frame detail away and the differences between DLSS modes start to show. Premium-tier streams (GeForce Now Ultimate at 75 Mbps, PS Plus Premium at 35-40 Mbps) are where the masking effect is strongest.

Path-traced content at high settings still shows mode-dependent quality on cloud — when there's enough detail in the source frame, the encoder can't compress it away cleanly. Cyberpunk 2077 Path Tracing on Ultra is the one case we tested where DLSS Quality vs Performance was visibly different on cloud. Other content was indistinguishable.

Cloud gaming services should default to aggressive DLSS/FSR modes (Performance or Balanced) rather than Quality. The frame rate benefit is real and the visual penalty is hidden by the video encoder. Most services currently default to Quality, which is the wrong choice for cloud and is leaving frame rate on the table.

GeForce Now in early 2025 quietly switched some titles to default DLSS Performance and the user-visible result was better, not worse. The change was unmarketed and most users didn't notice anything except smoother gameplay.

If your cloud service exposes DLSS/FSR settings in-game, choose Balanced or Performance modes rather than Quality. You'll get higher frame rates with no visible quality penalty in the cloud stream.

If you're toggling between local and cloud play on the same game, use different settings for each. Quality mode locally; Performance mode on cloud. The right answer depends on where the rendering and display chains converge, not on which mode the publisher's documentation recommends.