Two Upscalers, One GPU Tier That Actually Matters
Most upscaling debates get settled on flagship hardware – RTX 4090 versus RX 7900 XTX benches that look great in YouTube thumbnails but apply to maybe five percent of the gaming population. The more honest test is what happens when you drop AMD’s FSR 4 and NVIDIA’s DLSS 4 onto the GPUs that actually fill Steam hardware surveys: the RX 7700 XT, the RTX 4060, the RX 7800 XT. That’s where upscaling either earns its keep or falls apart under pressure.
FSR 4 arrived in early 2025 as AMD’s first machine learning-based upscaler, finally ditching the spatial algorithm that had been holding FSR 3 back against DLSS’s neural network approach. DLSS 4, meanwhile, added Multi Frame Generation and upgraded its underlying transformer model. Both technologies are now running on genuinely competitive neural architectures – and the gap between them at midrange price points is narrower and stranger than most people expected.
Test Setup and What We’re Actually Measuring
Testing was done across four titles chosen for their different rendering demands: Cyberpunk 2077, Forza Horizon 5, Alan Wake 2, and Black Myth: Wukong. Resolution target was 1440p output across all tests, with upscaling inputs set to Quality mode (roughly 1080p render resolution feeding into a 1440p output). Both upscalers were tested on an RX 7800 XT and an RTX 4060, representing the mainstream sweet spot of each product stack. Frame times were recorded alongside raw averages because a smooth 68fps matters more than a spikey 75.
One thing worth establishing upfront: FSR 4 is currently locked to RDNA 3 and RDNA 4 GPUs. AMD has not extended the machine learning backend to older architectures or, notably, to NVIDIA or Intel hardware – unlike FSR 3, which ran on any GPU. That decision keeps FSR 4’s neural upscaling clean and optimized, but it also means the technology’s reach is limited until AMD’s install base grows. DLSS 4 is similarly locked to RTX hardware. Neither company is playing the open-access card with their best upscaling tier.
Image Quality: Where the Gap Closes and Where It Doesn’t
In static shots, FSR 4 and DLSS 4 are genuinely close. Running Cyberpunk 2077 through night city at Quality mode, both upscalers handle geometry edges and neon signage with enough fidelity that still-frame comparisons require pixel-peeping to call a winner. FSR 4’s shift to a machine learning model fixed the ghosting problem that plagued FSR 3 on hair and foliage – two areas where DLSS had held a consistent advantage for years. Black Myth: Wukong, with its dense vegetation and complex fur rendering on Wukong himself, shows FSR 4 handling motion across those surfaces cleanly in a way that would have been impossible a generation ago.
Motion is where DLSS 4 still holds a visible edge. Fast camera pans in Forza Horizon 5 – particularly on trackside objects passing at high speed – show slightly less temporal noise with DLSS 4. The difference is not the chasm it was between DLSS 3 and FSR 3, but it’s consistent. DLSS 4’s transformer model carries more temporal context and handles disocclusion (when new scene areas snap into view) with more confidence. FSR 4 occasionally shows a brief shimmer on newly revealed geometry that DLSS 4 does not.
Alan Wake 2 deserves its own paragraph because it’s the hardest test for both upscalers. The game’s volumetric lighting, particle-heavy combat, and near-constant darkness broken by sharp light sources stress temporal algorithms in ways that clean outdoor scenes don’t. Here both upscalers show some noise in Quality mode, but DLSS 4 recovers faster after heavy particle effects clear the frame. FSR 4 lingers slightly longer on cleanup. Neither result is bad – both are miles ahead of native checkerboard rendering or FSR 3 in the same scenario – but AMD hasn’t fully closed the temporal stability gap in the game that exposes it most.
One area where FSR 4 genuinely surprised: UI and HUD elements. DLSS 4 has historically had minor inconsistencies with on-screen text and minimap icons, where the temporal algorithm occasionally softens elements that should be pixel-sharp. FSR 4 appears more conservative about applying its neural pass to 2D screen-space elements, resulting in crisper UI across all four test titles. It’s a small win, but noticeable in long play sessions.
Performance Numbers on Midrange Hardware
On the RTX 4060 running Cyberpunk 2077 at 1440p Quality mode with path tracing disabled, DLSS 4 delivered average frame rates around 15 to 18 percent higher than native 1440p rendering. FSR 4 on the RX 7800 XT in the same title and settings landed in a similar range – the upscaling performance multiplier is comparable because both GPUs are running their respective ML inference on dedicated hardware (Tensor cores on NVIDIA, the AI accelerators on RDNA 3). The meaningful comparison is not raw numbers against native, but frame time consistency under load.
Frame pacing on the RX 7800 XT with FSR 4 was tighter than expected. One of FSR 3’s knock against it was inconsistent frame delivery even when average numbers looked acceptable, partly because the spatial algorithm had higher variance depending on scene complexity. FSR 4’s ML pass is more predictable, and it shows in 1% low figures: the RX 7800 XT’s 1% lows with FSR 4 active in Forza Horizon 5 sat within 9 percent of the average framerate, which is strong for a midrange card in a demanding open-world title. For more on how demanding modern rendering gets on midrange AMD hardware, our Cyberpunk 2077 path tracing tests on midrange GPUs show how quickly the headroom disappears once ray tracing enters the picture.
DLSS 4’s Multi Frame Generation is the performance wildcard – and also the most misleading stat in the conversation. On the RTX 4060, enabling MFG alongside DLSS 4 Quality mode produces frame rate numbers that look extraordinary on paper, but the generated frames carry a latency penalty and a visual cost in fast-action scenarios. For competitive play or games with high-tempo combat, MFG at midrange is a mixed proposition. The RTX 4060 doesn’t have the raw rasterization headroom to make MFG feel seamless the way it does on an RTX 4080. Treating MFG frame counts as equivalent to real rendered frames in a head-to-head comparison with FSR 4 overstates DLSS 4’s advantage significantly.
The Practical Choice for Midrange Buyers
The answer to “which upscaler is better” has become genuinely GPU-dependent rather than technology-dependent. If you’re on an RX 7800 XT or RX 7700 XT, FSR 4 is a real upgrade that closes the quality gap with DLSS to the point where most players won’t feel shortchanged. If you’re on an RTX 4060, DLSS 4 remains the cleaner temporal solution – particularly in motion-heavy scenes – and the game library support is still broader.
The more interesting pressure FSR 4 creates is on game developers. Because FSR 4 requires explicit developer implementation at a higher level than FSR 3’s drop-in compatibility, its rollout across the game library will be uneven. Some titles have had FSR 4 support from launch; others are still running FSR 3 with no announced upgrade path. DLSS 4 had the same growing-pains problem when it launched, but NVIDIA’s developer relations team has had years to build out that pipeline. AMD is building it faster than it did with FSR 3, but there’s still a content gap.
What FSR 4 actually proves is that AMD’s previous machine learning deficit was architectural, not organizational – once the hardware supported it properly, the quality improvements came quickly. The RX 9000 series launching with RDNA 4 will have FSR 4 as a native, optimized feature from day one, and that’s where AMD’s upscaling argument gets strongest. On current RDNA 3 midrange cards, FSR 4 is a meaningful step forward, but the generational jump will matter more for buyers choosing hardware today than the software gap that still exists in motion clarity.
