AMD releases source code for FidelityFX Super Resolution 2.2 aka its DLSS

AMD FidelityFX Super Resolution 2 (FSR 2) hit version 2.2 late last year and has found its way into games like F1 22 and the latest Need For Speed. Since its creation, AMD has positioned FSR as an open-source alternative to NVIDIA's AI-powered DLSS. It is available in all games and hardware configurations - including consoles like the PlayStation 5 and Xbox Series X|S.

3

With most FSR 2 titles out there using older versions of the tech - including Cyberpunk 2077 - today comes the official open-source release of FSR version 2.2.0. As noted by AMD in the release, "FidelityFX FSR2 v2.2.0 is a major update that significantly increases image quality in many more cases," with the latest update "significantly reducing artefacts such as high velocity ghosting and shimmering."

Even though it sounds like an incremental update over FSR version 2.1.2, this latest version changes the FSR2 API, meaning that "updating to FSR 2.2 from 2.1.2 will take some work." However, AMD adds that it's still a straightforward implementation.

FidelityFX FSR2 v2.2.0 improves on FSR v2.1.2 in multiple ways, with the continued goal of increasing overall image quality and significantly reducing artefacts such as high-velocity ghosting and shimmering.

The application-side FSR2 API has changed, so updating to FSR 2.2 from 2.1.2 will take some work, but it should be straightforward. Developers should pay particular attention to the changes related to mask generation to get the best quality upscale.

You can grab AMD's FidelityFX FSR2 v2.2.0 via GitHub.

Here's a list of the changes.

• Introduction of debug API checker.
• Changes to improve "High Velocity Ghosting" situations.
• Changes to Luminance computation with pre-exposure application.
• Small motion vectors ignored in previous depth estimation.
• Changes to depth logic to improve disocclusion detection and avoid self-disocclusions.
• Dilated reactive mask logic updated to use temporal motion vector divergence to kill locks.
• New lock luminance resource.
• Accumulation overhauled to use temporal reactivity.
• Changed how intermediate signals are stored and tonemapped.
• Luminance instability logic improved.
• Tonemapping no longer applied during RCAS to retain more dynamic range.
• Fixes for multiple user reported issues on GitHub and elsewhere. Thank you for your feedback!