Introduction

DLSS 4 introduces a major leap in image quality and performance. It isn't just a version bump featuring Multi Frame Generation, but introduces updates to nearly all DLSS sub-features. DLSS from its very beginning relied on AI to reconstruct details in Super Resolution, and with DLSS 4, NVIDIA is introducing a new Transformer-based AI model to succeed the Convolutional Neural Networks previously used—for double the parameters, four times the compute performance, and significantly improved image quality. Ray Reconstruction, introduced with DLSS 3.5, also gets a big image quality update with the new Transformer-based model.



In this article we will look at the improvements the DLSS 4 Transformer model brings—specifically focusing on the improvements for owners of older GeForce RTX cards: GeForce 40 and older.


Transformer-based Super Resolution in DLSS 4 significantly enhances detail retention. Some examples include higher fidelity in textures, such as the bag's intricate patterns. Smarter models help improve image quality for heavily used features like Ray Reconstruction and Super Resolution.

Image Quality

Below, you will find comparison screenshots of Cyberpunk 2077, Alan Wake 2 and Stalker 2: Heart of Chornobyl at 4K, 1440p, and 1080p resolutions, showcasing different quality modes for DLSS Super Resolution CNN and Transformer models, FSR, and XeSS. Screenshots for TAA, DLAA, DLSS Frame Generation and DLSS Ray Reconstruction are also available in the dropdown menu. For a dynamic view of these technologies in motion, watch our side-by-side comparison video. This video reveals issues such as shimmering or temporal instability that might not be evident in still images.

All tests were conducted using a GeForce RTX 4080 GPU at Ultra graphics settings with Path Tracing enabled in Cyberpunk 2077 and Alan Wake 2. We also used a GeForce RTX 3060 and GeForce RTX 3080 for additional testing. Motion blur, depth of field, film grain and chromatic aberration were disabled for better image viewing.

Cyberpunk 2077 Screenshots

Alan Wake 2 Screenshots

S.T.A.L.K.E.R. 2: Heart of Chornobyl Screenshots

Side-by-Side Comparison Video

Conclusion

First of all, we'd like to congratulate NVIDIA on their new release, bringing impressive image quality improvements for DLSS upscaling and Ray Reconstruction not only for the latest generation of RTX cards, but also supporting even the six-year-old GeForce RTX 20-series GPUs. Although the performance penalty for running a more advanced AI model on something like an RTX 2060 could be challenging in a few specific scenarios (more on that will be in our performance evaluation section), but the ability to run DLSS upscaling and Ray Reconstruction in its latest and greatest form on any RTX GPU is a big win for everyone.

Super Resolution Upscaling
According to NVIDIA, DLSS Super Resolution with Transformer model provides improvements to temporal stability, less ghosting, higher detail in motion, alongside several image quality related bugs and glitches that were fixed using the new model. To evaluate these claims, we've tested the new model in three completely different games that were built on different engines, to collect more accurate data:

• In Cyberpunk 2077, a significant reduction in ghosting is the most noticeable upgrade, and it is especially visible on animated textures. The previous CNN model for Super Resolution always had issues in this aspect, for example, displaying moving text on TV or LED screens in general, with excessive ghosting and smearing. This happened because motion vectors from the game, that are describing how things are moving around, don't go along with the textures. To avoid this, the new Transformer model was trained to deal with this specific issue by analyzing the scene and recognizing animated textures to properly display what's on them—and the results are incredible. All moving text on animated textures is now perfectly stable and readable from a distance in all quality modes and resolutions, making it basically a night and day difference compared to the CNN model. The Transformer model also fixes one specific bug that is related to Ray Tracing or Path Tracing: the random bloom flashes in daytime when sunlight is reflected on moving vehicles—this issue is gone with the new model.
• In Alan Wake 2, the biggest upgrade is that the shimmering in vegetation is significantly reduced with the new DLSS model, especially in "Performance" mode across all resolutions. Speaking of "Performance" mode, this quality mode received a significant upgrade in terms of overall image clarity across all resolutions, making it noticeably less blurry and more stable in motion. The best example is to look at our screenshots and compare CNN and Transformer models at 1440p "Performance" mode, and the difference in details is very obvious, especially on small thin objects such as power lines. Of course, screenshots don't tell the whole story, but using it in person for at least an hour, we must say that this low resolution quality mode is very usable and playable with the new model. It's actually impressive how the new model manages to produce an image this good, considering the fact that the internal resolution is 720p. Just to be clear, the improvements to "Performance" mode are not related to Alan Wake's implementation only, but rather to the Transformer model as a whole, which means you will see similar improvements in other games as well.
• The image quality and overall stability in S.T.A.L.K.E.R. 2: Heart of Chornobyl is also significantly improved when using the new Transformer model. The shimmering in vegetation is significantly lowered, similarly to what we've seen in Alan Wake 2, especially at lower resolutions and lower quality modes. Another very noticeable improvement is how the new model deals with the stability of thin objects, such as power lines for example. Even in our screenshots it is very noticeable that the power lines in the TAA image are completely non-existent. While DLSS Super Resolution with the CNN model struggles to keep details stable, DLSS Super Resolution with the Transformer model is producing a perfectly stable image in comparison—this aspect is better highlighted in our side-by-side video as well. Also, in the same video, take a closer look at laser sights in motion: in the CNN model the stability of the laser sights is very poor, rendering them with ghosting and smearing, while the Transformer model produces a significantly more stable image.

Ray Reconstruction
Out of all image quality upgrades in the DLSS 4 package, Ray Reconstruction received the most impressive improvements:

• The Ray Reconstruction implementation in Cyberpunk 2077 is known for being quite problematic in terms of creating an excessive "oil painting" effect on the whole image, ghosting and smearing artifacts on distant objects or while driving, and a somewhat washed out rendering of the character faces that degraded detail, especially in motion. Although, even with the CNN model, Ray Reconstruction had a lot of benefits and improvements too, such as more stable and detailed shadow rendering or significantly improved resolution of ray traced reflections. Some players complained that Ray Reconstruction introduced immersion-ruining artifacts for some players. When we first booted up the game using Ray Reconstruction with the Transformer model, we instantly saw a huge improvement—the oil painting effect is reduced to a minimum on the whole image across all resolutions. The ghosting and smearing artifacts in low contrast scenarios are gone as well, producing a very stable image in motion. Faces of characters are significantly cleaned up, they're very stable in motion and do not affect immersion with distracting artifacts during cutscenes or dialogues like with the old CNN model. Animated textures received the same treatment as with DLSS Super Resolution, eliminating distracting ghosting and smearing artifacts across all resolutions.
• Overall, the new model for Ray Reconstruction greatly enhances the quality of RT effects, producing improved details in RT reflections across all resolutions, especially in motion, which results in increased sharpness of some reflective materials, such as the road after rain, for example. The quality of RT shadows is also affected, changing the appearance of RT shadows at close range and improving detail, stability and sharpness at far distances. The direct and indirect lighting with Ray Reconstruction becomes more responsive, and the quality of lighting is noticeably more accurate and realistic, greatly improving the overall experience.

Frame Generation
In terms of overall image quality, the new model for Frame Generation is the least impressive upgrade compared to DLSS Super Resolution and Ray Reconstruction. Transformer Frame Generation has a less blurry image in motion and less ghosting, but those improvements mostly come from the Transformer model Super Resolution image as it is used as a base for Frame Generation. Instead of image quality, the new model for Frame Generation is focused on improving other aspects: increased framerates and smoothness, with less VRAM usage at the same time.

• The framerate increase will depend on the game, output resolution and only if you're GPU limited: compared to the old CNN model, the Transformer model can generate up to 15% more frames, further improving the smoothness in games like Alan Wake 2 or Stalker 2: Heart of Chornobyl at 4K when the CNN and Transformer models are compared directly, while at 1440p and 1080p resolutions the uplift in framerate is usually below 10%. We also tried a few additional games to collect more performance data, and in Horizon Forbidden West we saw a 15% FPS improvement at 4K resolution, while in The Witcher 3 the uplift was only at around 5% at 4K.
• VRAM usage improvements will also depend on a game and your output resolution: at 4K resolution you can expect up to 600 MB less VRAM usage, up to 300 MB at 1440p and up to 200 MB at 1080p when DLSS Frame Generation is enabled.

Performance
The new Transformer model for DLSS Super Resolution does have a performance cost compared to the previous CNN model. On a GeForce RTX 4080 GPU, the performance cost is very small, usually around 3-5% depending on a game, and enabling Ray Reconstruction in conjunction with Super Resolution has a similar performance boost as with the CNN model, which is good. However, the performance cost from Transformer model Super Resolution and Ray Reconstruction together is different on GeForce RTX 30 series GPUs. As an example, we've tried running the new model in Cyberpunk 2077 on an RTX 3060, one of the most popular GPUs currently, and here is the result of what we've discovered:

• With native TAA at 1080p and "Ultra" RT preset the game runs at 25 FPS (base performance).
• Enabling DLSS SR CNN and "Ultra" RT preset will lead to a 45% performance boost (44 FPS).
• Enabling DLSS SR CNN with RR CNN and "Ultra" RT preset will lead to another 14% performance boost (51 FPS).
• Enabling DLSS SR Transformer and "Ultra" RT preset will lead to a 5% performance loss compared to DLSS SR CNN (42 FPS).
• Enabling DLSS SR Transformer with RR Transformer and "Ultra" RT preset will lead to another 7% performance loss compared to DLSS Transformer (39 FPS), and a massive 25% performance loss compared to DLSS SR + RR CNN.

To sum it up, while the Transformer model for Super Resolution runs at very minor performance cost of a 5% on an RTX 3060, running Transformer Super Resolution and Ray Reconstruction at the same time will significantly drop your performance by 25% compared to DLSS SR + RR on the CNN model, which is quite massive. Some could argue that an RTX 3060 is struggling with ray tracing anyway, thus Ray Reconstruction isn't really a thing on this GPU, but we also tried to run the same test on an RTX 3080, a more powerful Ampere GPU and very capable of running ray tracing—and the results were the same, indicating that the quality of Ampere's tensor cores is the issue, not the quantity.