Something to look forward to: Unlike Nvidia's DLSS and Intel's XeSS, AMD's FSR upscaling technology does not rely on AI or machine learning. Instead, it uses a combination of spatial and temporal upscaling, anti-aliasing, and other techniques to enhance image quality from lower-resolution inputs. While AMD's solution supports a wider range of hardware than its competitors, it typically delivers lower image quality. However, AMD has now confirmed that it is moving forward and has been working toward incorporating AI for some time.

Jack Huynh, senior vice president of AMD's Computing and Graphics Business Group, recently announced that FidelityFX Super Resolution 4.0 (FSR 4) will use AI. This marks a significant departure from previous versions but potentially opens more questions than the confirmation answers.

Competing upscaling solutions like Nvidia DLSS and Intel XeSS use AI and machine learning to upscale images while minimizing performance impact, though they require specialized hardware not found in game consoles or older GPUs.

In contrast, the latest version of FSR, FSR 3.1, uses spatial and temporal techniques for upscaling and frame generation. It works on nearly any device, but benchmarks show it produces less stable images compared to AI-driven solutions.

Sony's introduction of AI-based PlayStation Spectral Super Resolution as a key feature of the PlayStation 5 Pro signals a broader industry shift toward AI upscaling – not to mention, the console is based on AMD's graphics tech. Additionally, the upcoming successor to the Nintendo Switch is expected to support DLSS.

At IFA 2024, Huynh explained that AMD had focused on analytical and filter-based techniques with FSR 2 and 3 to accelerate development, but an AI-based solution has been in the works for around a year.

Interestingly, Huynh's comments on FSR 4 centered around maximizing battery life on handheld gaming PCs through frame generation. The timing suggests that FSR 4 might be optimized for the AMD Z2 Extreme mobile processor, which the company plans to release early next year. The SoC's next-generation NPU could be the cornerstone of AMD's AI upscaling technology.

Huynh did not mention whether FSR 4 would be available on more traditional products like RDNA 4 graphics cards, which are expected to debut in 2025 (though it would only make sense for that to be the case). It also remains unclear whether older AMD GPUs will support the new upscaling technology.

While AMD typically allows its software to work across a wider range of hardware, including competitors' devices, all AI-based upscaling solutions so far have required some form of dedicated hardware.

Permalink to story:

AMD FSR 4 will transition to pure AI upscaling, following DLSS and XeSS

They just need to add a small network to the end of the FSR pipeline that receives several characteristics of the input image. In the patent they published a few years ago this network appears, but more simplified. I modeled it in PyTorch and did some experiments training it with a dataset I had made some time before with a couple of million patches extracted from images from The Witcher III, and for being small it delivered an acceptable result, of course, it was unstable because it did not use temporal information, or depth or movement, or anything else, just color. There are some networks for SR on mobile devices, very simple and of better quality, besides being fast and consuming little energy. And they can use some of this or derivatives.