Most of the misunderstandings around image-upscaling technologies stem from how they were born. The first generation still used data from a single frame to rescale to a higher resolution. Since then, explanations of super sampling or super resolution have often repeated a simplified line: the game engine renders at a lower resolution and artificial intelligence fills in the rest. It supposedly draws on training from the same games at higher resolution and, because it \u201cknows\u201d how the frames should look in higher quality, it \u201cimagines\u201d the missing details into the low-resolution frame.<\/p>\n
The popularity of this explanation is reinforced by AI models for enlarging photos, which today can add details to old, low-quality pictures that were never there. Sometimes they miss, creating a wrong shape or inserting a pattern or color that doesn\u2019t belong. That\u2019s why gamers worry that AI upscaling in games will add things that shouldn\u2019t be there\u2014things that wouldn\u2019t appear with proper native rendering.<\/p>\n
We\u2019ll examine in one of the following chapters why, if super resolution worked that way, it could not achieve the level of detail that modern reconstruction technologies do today.<\/p>\n
Trust isn\u2019t helped by the technology developers and PR teams themselves, who talk about each new generation of upscaling in superlatives.<\/p>\n
They try to ride the AI wave, attribute miraculous abilities, and sometimes oversimplify. Only when they present the next generation\u2019s improvements do they admit the shortcomings of the previous one.<\/p>\n Another group of opponents argues that with the advent of upscaling, developers neglect graphics optimization\u2014rather than tuning visuals, they push gamers to use upscaling, which will ultimately lead to lower graphics quality.<\/p>\n But upscaling as such won\u2019t hurt visual quality\u2014if anything, the opposite. From the very beginning, game artists have wrestled with how to work around technological limits. Naturally, they would like to render everything as faithfully and as high-quality as possible, but there has never been, is not, and for a long time will not be enough performance for that.<\/p>\n Developers have several ways to deal with limited performance. The simplest is to restrict players\u2019 graphics settings. If they cap the maximum detail level at a point where the game runs smoothly on most cards, they can count on praise for good optimization. But the downside is that a year or two later, when more powerful GPUs arrive, you still won\u2019t be able to unlock higher details in that same game. At launch it may run smoothly but look average\u2014and its visuals will age faster.<\/p>\n Some developers go the opposite route, deliberately including settings that current hardware can\u2019t handle at maximum quality. They bank on more powerful GPUs arriving later, so they add \u201cexperimental\u201d options. But then they often get criticized for poor optimization because the game isn\u2019t playable at maximum settings even on the fastest hardware available.<\/p>\n For developers taking that approach, image-upscaling techniques are a gift. And players see it that way too\u2014especially those who want to enjoy new games at high detail levels without owning top-tier, latest-generation GPUs.<\/p>\n Upscaling shouldn\u2019t be dismissed as a patch for hardware or software shortcomings. It\u2019s more a way to push graphics quality higher, even when raw computational performance isn\u2019t enough to \u201cfaithfully render every pixel.\u201d<\/p>\n Faithful rendering means redrawing the entire frame, point by point, up to a hundred times per second\u2014even when very little has changed between frames. That\u2019s a huge waste of computational capacity, which could be put to better use.<\/p>\n In the next chapters, we\u2019ll look at how upscaling used to work, and how things improved with the advent of temporal supersampling\u2014using samples gathered over time from multiple frames. We\u2019ll demonstrate this with the popular Grand Theft Auto V Enhanced<\/strong>, a remaster of the original game where developers added temporal supersampling support alongside ray tracing.<\/p>\n For testing and comparison, we\u2019re using the budget-friendly Radeon RX 9060 XT Gaming OC 16G<\/a> from Gigabyte, which I recently reviewed on HWCooling<\/a>.<\/p>\n To remind you of the key specifications of the card:<\/p>\n You can find a more detailed description of the card in the review and on the Gigabyte product page<\/a>.<\/p>\n We\u2019ll demonstrate how one of AMD\u2019s technologies, Anti-Lag, works in practice using a top-tier gaming monitor, the AORUS FO27Q3, also from Gigabyte<\/a>. It uses a QD-OLED panel developed by Samsung. With a resolution of 2560 \u00d7 1440 pixels, it offers a 360 Hz refresh rate and is certified with VESA ClearMR 13000 and VESA DisplayHDR True Black 400. For gamers, the major advantage over LCD monitors is that OLED can redraw pixels with significantly faster response times.<\/p>\n The basic monitor specifications are summarized in bullet form below. For a detailed description and demonstrations of supported technologies, visit the monitor\u2019s product page<\/a>.<\/p>\n Image \u201cupscaling\u201d technologies have already won over most gamers. But there is still a camp of staunch opponents who prefer a faithfully rendered image over one \u201cmade up\u201d by artificial intelligence. On a Radeon RX 9060 XT with AMD FSR 4, we\u2019ll look at why that reading is inaccurate and how the new algorithms can turn low-resolution frames into images with a level of detail comparable to honest native-resolution rendering.<\/p>\n The following comparison shows where the main issue lies when upscaling from a low resolution. We\u2019ll examine how many details are retained in native resolution versus when rendering at half the resolution. Below are three images with 1:1 crops that can be enlarged to double size when clicked.<\/p>\n The first crop shows the game rendered at its native resolution of 2560 \u00d7 1440 pixels. The next examples are rendered at half the resolution\u20141280 \u00d7 720 pixels (or one-quarter of the original area)\u2014and then upscaled using different methods, including AI with a trained neural network. A resolution of 1280 \u00d7 720 should roughly match<\/strong> what FSR uses when your monitor\u2019s native resolution is 2560 \u00d7 1440 and you enable FSR in Performance mode<\/strong>.<\/p>\n The second crop shows an image rendered at a resolution of 1280 \u00d7 720 pixels. This renders an area four times smaller with four times fewer pixels. To match the scale, I enlarged it using a sharp nearest-neighbor filter. Each rendered pixel is thus stretched to 2\u00d72 pixels. The resulting size then corresponds to a native resolution image.<\/p>\n At 1280 \u00d7 720, a lot of detail is already lost. Small text becomes unreadable because each letter only gets a handful of pixels. When you interpolate this to a higher resolution, the image just gets blurred to hide jagged edges, but the small letters will remain illegible.<\/p>\n The third crop demonstrates how the game\u2019s built-in upscaling handles this situation. The frame scaler is set to 1\/2 (x0.500). Jagged edges are less noticeable now, but overall sharpness drops as a result. The fourth crop shows what happens when you set the game to 1280 \u00d7 720 and let the hardware scaler on the GPU or monitor upscale it. In this case, it was likely the scaler on an AverMedia GC573 capture card, which scaled the image back to 2560 \u00d7 1440.<\/p>\n The last crop illustrates what happens when a neural network does the upscaling. Here, the image was upscaled by AI using the Upscayl application<\/strong>, which employs models like Real-ESRGAN. The upscaling was done in Standard mode.<\/p>\n Even if I overlook the fact that the text remains unreadable, this implementation cannot be used in games. Upscaling a single frame takes on common PC over five seconds.<\/p>\n It’s clear that text is only legible at the native 1440p resolution. In this crop, you can see that AI does a nice job smoothing edges, enhancing recognized shapes, and sharpening the image. However, it can’t always reconstruct the original shape from just a few pixels. This is particularly evident with the stars on the red poster or the readability of text. From a four-pixel character, the AI cannot determine what letter was originally there; it replaces text with geometric shapes, leaving it unreadable. And I should remind you that upscaling a single frame to this quality takes about five seconds using GPGPU.<\/p>\n Although simple interpolation performed in-game and on the scaler don\u2019t differ fundamentally in terms of rendering, the in-game method does have one advantage: while the image is rendered at a lower resolution, additional graphical interface elements like text or HUD are drawn over it at the display\u2019s native resolution. This is visible in the lower left corner of the frames with the map\u2014with in-game upscaling, the overlay map and text retain the same quality as in native resolution, whereas with all other interpolation methods, the interface is enlarged and becomes jagged, blurry, or harder to read.<\/p>\n Below are the full JPEG images with 99% compression quality. Remember, it\u2019s essential to view them at a scale where one image pixel corresponds to one screen pixel. If you downscale or upscale them, interpolation in the browser will blur them, and you won\u2019t see the differences at a significant downscaling.<\/p>\n native 1440p<\/a> | native 720p<\/a> | in-game scaling 0,5<\/a> | hardware scaler<\/a> | Upscayl (AI)<\/a><\/p>\n ![\"\"](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/08\/amd-fsr-4_005.jpg\")
<\/a><\/p>\nWhen you want more than hardware can handle<\/h2>\n
<\/a><\/p>\n
\n<\/a><\/a><\/a><\/a><\/a><\/p>\nRadeon RX 9060 XT: The Most Affordable Addition to the New Radeon Family<\/h2>\n
\n
<\/a><\/p>\n<\/a><\/p>\n\n
When the image is really upscaled by AI<\/h2>\n
\n<\/a><\/p>\n<\/a><\/p>\n<\/a><\/p>\n<\/a> <\/a> <\/a> <\/a> <\/a><\/p>\n
![native 1440p<\/a> |](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/08\/gtav-rtx-max-010-native-1440.jpg\"){kind=link}
![native 720p<\/a> |](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/08\/gtav-rtx-max-011-native720p-05.jpg\"){kind=link}
![in-game scaling 0,5<\/a> |](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/08\/gtav-rtx-max-012-ingame-scaler-05.jpg\"){kind=link}
![hardware scaler<\/a> |](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/08\/gtav-rtx-max-013-gpu-scaler-05.jpg\"){kind=link}
![Upscayl (AI)<\/a><\/p>\n \r\n\r\n \r\n](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/08\/gtav-rtx-max-014-native720-upscayl-200.jpg\"){kind=link}