FSR (FidelityFX Super Resolution) is AMD’s family of upscaling technologies that render a game at a lower internal resolution and then reconstruct an image at the display resolution, raising frame rates while keeping the picture close to native quality. Among graphics settings, FSR is one of the main tools for trading a small amount of image quality for a large gain in performance.
Unlike vendor-locked approaches, most FSR versions run on graphics cards from AMD, NVIDIA and Intel as well as on consoles, which has made it a common option in PC games and a frequent choice in console development. The family now spans spatial upscaling, temporal upscaling, frame generation and, in its newest generation, machine learning.
How FSR reconstructs the image
FSR works by rendering fewer pixels and filling in the missing detail. The game draws each frame at, for example, 1440p for a 4K output, and FSR scales it up before the user interface is drawn at full resolution. The quality of the result depends on how much information the algorithm can use beyond the single low-resolution frame.
FSR 1 is a spatial upscaler: it analyzes only the current frame, using an edge-adaptive scaling pass followed by a sharpening pass. It is cheap and easy to integrate but cannot recover detail that the low-resolution frame never contained, so thin lines and distant textures suffer.
FSR 2 and later are temporal upscalers. They combine samples from several previous frames, using motion vectors and depth data from the engine to align old pixels with new ones, and apply camera jitter so each frame samples slightly different positions. Accumulated over time, these samples recover detail close to or occasionally beyond native rendering, and the upscaler also replaces the game’s own temporal anti-aliasing.
FSR versions and what changed
Each FSR generation added a different capability, and the version a game ships with has a larger effect on image quality than the quality preset. The table summarizes the main generations.
| Version | Method | Key characteristics |
|---|---|---|
| FSR 1 | Spatial upscaling and sharpening | Very light, works almost anywhere, weakest detail |
| FSR 2 | Temporal upscaling without machine learning | Needs motion vectors, much better detail, some ghosting |
| FSR 3 | Temporal upscaling plus frame generation | Adds interpolated frames, requires low-latency support |
| FSR 3.1 | Decoupled upscaler and frame generation | Upgradable through the driver, works with other upscalers |
| FSR 4 | Machine learning upscaling | Best stability, targets newer Radeon hardware |
FSR 1 to 3.1 were released as open source in the FidelityFX SDK, which encouraged adoption across engines and platforms. The machine learning generation relies on dedicated AI acceleration in recent Radeon GPUs, so its availability is narrower than earlier versions, a trade-off similar to that of competing ML upscalers.
Quality modes and render resolutions
FSR exposes quality modes that set the scaling factor between internal and output resolution. A higher factor means fewer rendered pixels, more frames per second and more reconstruction work for the algorithm.
| Mode | Scale factor per axis | Internal resolution at 4K |
|---|---|---|
| Native AA | 1.0x | 3840 x 2160 |
| Quality | 1.5x | 2560 x 1440 |
| Balanced | 1.7x | About 2259 x 1270 |
| Performance | 2.0x | 1920 x 1080 |
| Ultra Performance | 3.0x | 1280 x 720 |
Performance gains are not proportional to the pixel count. A frame that is limited by the processor, by fixed post-processing costs or by the upscaler pass itself gains less. In practice, Quality mode often brings a gain of roughly 30 to 60 percent in GPU-limited scenes, but the figure varies widely with the game and hardware. Higher output resolutions tolerate stronger modes better, because 4K Performance still renders 1080p, while 1080p Performance renders only 540p.
FSR compared with DLSS and XeSS
FSR competes with NVIDIA DLSS and Intel XeSS, which follow the same principle of temporal reconstruction. The practical differences lie in hardware requirements, the use of trained neural networks and the resulting stability of fine detail in motion.
| Technology | Hardware support | Reconstruction approach |
|---|---|---|
| FSR 2 and 3 | Most modern GPUs and consoles | Hand-tuned temporal algorithm |
| FSR 4 | Recent Radeon GPUs | Trained neural network |
| DLSS | NVIDIA RTX GPUs only | Trained neural network on Tensor cores |
| XeSS | Intel Arc natively, other GPUs via fallback | Neural network, faster path on Intel |
In side-by-side comparisons, non-ML FSR has typically shown more shimmering, disocclusion artifacts and ghosting than DLSS, especially at lower resolutions, while the ML-based generation has narrowed that gap considerably. The main strength of the open versions remains broad compatibility, including handheld PCs and consoles.
Frame generation and latency
FSR frame generation inserts an interpolated frame between two rendered frames, using motion vectors and optical flow. It can nearly double the displayed frame rate, but the inserted frames do not respond to input, so input latency does not improve and actually increases slightly because one real frame must be held back.
For this reason, frame generation works best when the base frame rate is already around 60 fps and is paired with a latency reduction feature such as AMD Anti-Lag. Used on a base of 30 fps, it produces smooth-looking motion that still feels sluggish to control and shows more artifacts on fast-moving user interface elements.
Typical artifacts and how to reduce them
- Ghosting: trails behind moving objects when old samples are reused incorrectly; often worse on particles without proper motion vectors.
- Shimmering: flickering on thin geometry such as fences and foliage; a stronger quality mode or a newer FSR version usually helps.
- Disocclusion noise: blurry or noisy patches where hidden areas suddenly appear; it is most visible at low internal resolutions.
- Oversharpening: halos around edges from an aggressive sharpening slider; lowering sharpness restores a natural look.
- Mismatched user interface: blurred text or markers if a game upscales the interface together with the scene.
The idea in brief
FSR lets a graphics card draw a smaller picture and enlarge it intelligently. Drawing fewer pixels is much faster, and the enlargement step uses clues from previous frames to fill in detail, so the result looks nearly as sharp as the full-size picture.
The trade-off is simple: more frames per second in exchange for occasional small flaws, such as slight shimmer on thin objects. The higher the screen resolution, the less noticeable those flaws usually are.
A practical scenario
A player with a mid-range graphics card runs a demanding action game on a 4K television and reaches only about 35 frames per second at native resolution. Switching FSR to Quality mode makes the game render internally at 1440p, and the frame rate rises into the high 40s with an image that remains sharp from couch distance.
The player then tries Performance mode, which reaches a steady 60 fps but shows noticeable shimmer on distant foliage. As a compromise, they keep Quality mode and enable frame generation together with Anti-Lag, which delivers smooth motion on the television while camera control still feels acceptable.