NVIDIA's DLSS 4 Exits Beta: Transformer Model Revolutionizes AI-Powered Upscaling

DLSS 4 Out of Beta: Nvidia Transformer Model Enhances Image Quality

Nvidia's latest update to its AI-powered upscaling technology, DLSS 4, has officially moved out of beta with the release of a new Transformer-based model. DLSS (Deep Learning Super Sampling) has been a key technology in improving gaming visuals and performance by rendering games at lower resolution and then upscaling them using AI. With DLSS 4, Nvidia takes a big step forward in image quality and efficiency compared to its previous versions. In the earlier DLSS versions, Nvidia relied on Convolutional Neural Networks (CNNs) to perform image upscaling. CNNs work by analyzing clusters of neighboring pixels in each frame and estimating what new pixels to generate based on this localized information. This technique allowed for fast real-time upscaling but came with some drawbacks. Because CNNs rely on approximations over pixel groups, they sometimes introduced visual artifacts -- imperfections where the AI's predictions were slightly off. As games and display resolutions advanced, these limitations became more noticeable, prompting Nvidia to look for a better method. Enter the Transformer model, which Nvidia adopted for DLSS 4. Unlike CNNs, Transformers don't just focus on small groups of pixels; instead, they analyze every single pixel in the frame to assess its importance and contribution to the final image. This approach means the AI can reconstruct much finer details with higher precision, avoiding the kind of blurring or artifacting that CNNs sometimes produced. The result is an upscaled image that is much sharper and closer to the quality of native resolution rendering. This detailed pixel-by-pixel analysis requires more advanced processing but Nvidia has optimized DLSS 4 to maintain real-time performance. The Transformer model balances image quality improvements with rendering efficiency, allowing gamers to enjoy better visuals without sacrificing frame rates. Nvidia's move to this architecture also aligns with broader trends in AI, where Transformer networks have proven highly effective in tasks involving complex pattern recognition. DLSS 4's improvements come at a time when competing upscaling technologies, such as AMD's FidelityFX Super Resolution (FSR) 4, are also evolving. However, Nvidia's continued investment in AI-driven upscaling with DLSS keeps it positioned as a leading solution for gaming image enhancement. The official release of the Transformer model signals that Nvidia is ready to fully integrate this advanced technology into its software ecosystem. For gamers, this means sharper images, less visual noise, and better overall immersion when playing titles that support DLSS 4. It also opens the door for developers to push games with higher detail settings and resolutions while maintaining smooth performance. The transition from CNNs to Transformers is a technical milestone that highlights the role of AI in the future of real-time graphics rendering.

DLSS 4's impressive Transformer model for Super Resolution exits Beta

DLSS 4's upgraded Transformer-based AI Super Resolution has exited Beta, which means we'll start seeing it arrive in a lot more games. As an Amazon Associate, we earn from qualifying purchases. TweakTown may also earn commissions from other affiliate partners at no extra cost to you. NVIDIA announced the arrival of DLSS 4 alongside its big GeForce RTX 50 Series reveal at CES 2025. One of the most significant advances that DLSS 4 brings is a shift to a new and advanced Transformer model for Super Resolution AI upscaling, replacing the older and existing CNN, or Convolutional Neural Network, model. At a high level, the new Transformer model can examine every pixel in a rendered frame instead of a partial solution that makes estimates, resulting in a more detailed, sharper, and cleaner image comparable to a natively rendered image. With double the parameters, the Transformer model is also faster than CNN, reducing the latency from 3.25ms to just 1ms on a GeForce RTX 5090. DLSS 4 has been a game changer for image quality. Using the Performance preset in 4K delivers better image quality than the previous version's Quality preset. DLSS 4's Transformer model is also used in its Ray Reconstruction technology, which dramatically improves the quality of ray-traced effects (as seen above). At the time of the GeForce RTX 50 Series launch, DLSS 4 and the new Transformer-based Super Resolution were available in Beta form in over 70 games. That figure now sits in the 120+ region thanks to NVIDIA's inclusion of a driver-based DLSS Override feature as part of the NVIDIA App. And now, with the arrival of NVIDIA DLSS SDK 310.3.0, DLSS 4's upgraded Super Resolution has exited Beta. What does this mean? We'll start seeing many more games adding native support for DLSS 4 and the new Transformer-based Super Resolution. This can already be seen in games like Diablo 4, which is adding native DLSS 4 support next week as part of the game's big Season 9 update. This means you no longer have to use the DLSS Override feature to access DLSS 4; it'll be in-game as soon as you enable DLSS. With DLSS 4 exiting Beta, it also means that new releases supporting the older CNN model will become a thing of the past as DLSS 4's new and improved Super Resolution is compatible with all GeForce RTX graphics cards - from the GeForce RTX 2060 right up to the recently announced GeForce RTX 5050.