BenchmarksNews

NVIDIA DLSS 4 vs DLSS 3 Image Quality Compared: Years Ahead of AMD FSR 3 & Sony PSSR

NVIDIA announced DLSS 4 alongside the GeForce RTX 50 series “Blackwell” GPUs at CES earlier this month. Multi Frame Generation (MFG) is exclusive to the RTX 5090 and its siblings, but improved upscaling is coming to all RTX GPUs. The shift from a CNN to a Transformer-based neural network delivers sharper images, retaining more detail and ray-traced lighting samples across frames. The upgraded frame generation model is faster and uses 30% less VRAM than existing implementations.

Further Reading:

DLSS 4 vs. DLSS 3: Transformer vs. CNN

DLSS 3 and 2 leverage Convolutional Neural Networks (CNNs) which work by identifying patterns (or patches) across images, including edges, shapes, colors, textures, depth, etc. Feature maps are formed for each pattern and by learning from one another, they form complex maps capable of recognizing intricate geometry. Motion vectors help track these patterns across frames, and the high-resolution frames are constructed by comparing the low-resolution input to the previously rendered high-resolution frame.

DLSS 4 replaces CNNs with a transformer-based model capable of self-attention operations. This allows the model to determine the importance of each pixel (or pattern) across multiple frames. The resulting image is more stable and detailed with reduced ghosting and aliasing in motion. In addition to improving the upscaling quality, the transformer model produces superior ray-traced output, especially concerning reflections and diffuse lighting.

DLSS 4 vs. DLSS 3 Image Comparisons

We tested DLSS 4 and its transformer model in Cyberpunk 2077, Alan Wake 2, and Hogwarts Legacy. There are substantial image quality gains compared to DLSS 3.7 mainly with thin, intricate geometry, diffuse shadows, and reflections. All the below comparisons were made at 4K with path tracing, and DLSS set to “Performance Mode.” Click on an image to enlarge it.

DLSS 4 vs. DLSS 3: Shadows & Geometry Detail

The difference between the above two shots is clear as day. DLSS 4’s transformer model is much better at retaining complex geometric and lighting detail. The most obvious is the retention of the pinkish hue, characteristic of sunlit evenings. The ambient shadows along crevices, boundaries, and recesses are also more detailed.

The above 2x closeup shows how much more accurate DLSS 4’s transformer model is with thin, elaborate objects like wire fences. The textures are more detailed (not blurry), accurately casting shadows in the background.

In comparison, FSR 3 and XeSS 2 are a complete mess, losing much detail near the center and missing surrounding shadows and ambient occlusion.

The above comparison highlights the massive difference in ambient occlusion or diffuse shadow quality of the two upscaling models. DLSS 4’s transformer model accurately renders ambient shadows in bright daylight, while DLSS 3 loses most of them, likely washed out by indirect lighting. Likewise, FSR 3 and XeSS 2 are missing much of the shadowing, producing aliased fencing. DLSS 4 subtly retains the shadows on the wiring otherwise missed by it’s peers.

The below comparison highlights the improved detail retention along the bridge base and the significant reduction in blurring/smudging with DLSS 4. The palm trees are shadowed more uniformly, casting denser shadows near the center and more illuminated at the top.

The images upscaled by FSR 3 and XeSS 2 are quite noisy due to oversharpening and lose a ton of color and shadowing throughout the scene.

DLSS 4 vs. DLSS 3: Lighting & Reflections

DLSS 4 and its transformer model do wonders for specular lighting or reflections. Not only are the reflections clearer, but they retain more detail pertaining to distant objects, with specular surfaces accurately scattering light in the scene. This in-turn enhances the global illumination, producing accurate shadows and detailed reflections (see the planters).

FSR 3 and XeSS produce blurry reflections but aren’t much worse than DLSS 3.7 in other aspects. The specular surfaces are handled similarly.

Here’s another example of how DLSS 4 enhances specular reflections and lighting over DLSS 3.7, and FSR 3. Reduced noise, distant object coverage, and superior color absorption are worth noting.

DLSS 4 vs. DLSS 3: Alan Wake 2

Alan Wake 2 leverages FSR 2 rather than the newer implementation. Consequently, we’ll limit our comparison to DLSS 3 and 4 with the graphics settings maxed out and the upscaling preset set to performance.

Enlarging the image reveals some key differences. First off, DLSS 4 renders longer, more complete reflections of the colorful lighting prevalent throughout the scene. The light is thoroughly scattered along enclosed and distant surfaces, with light sources further off also contributing to the nearby (diffuse) lighting. Lastly, the raindrops (top-right) aren’t retained using DLSS 3. DLSS 4 doesn’t make the same error.

The above is a good example of how DLSS 4 renders more accurate diffuse “softshadows, while DLSS 3 results in blurry silhouettes with noise around the boundaries. Notice the shadows cast by the sofa, pillar, and the desk. The transformer model removes shadow noise and retains the diffuse lighting illuminating the inter-tile recesses. Strangely, it’s also blurring the tile patterning.

DLSS 4 vs. DLSS 3: Hogwarts Legacy

Hogwarts Legacy only features ray-traced shadows, reflections, and ambient occlusion. Somehow, it still runs worse than Cyberpunk 2077 with path tracing. The below example shows how DLSS 4 improves detail retention along the finer ends of the tree branches.

While the differences are relatively subtle, you can see that DLSS 4 retains more detail than native 4K. You’re rendering a higher quality 4K image by taking in a 1080p input at nearly twice the framerate.

The below is an excellent example of how DLSS 3 loses some of the diffuse shadows visible at native resolution (see the roofs). DLSS 4 not only retains those shadows but one-ups native by improving lighting detail on and around the lamp post.

The below comparison from the “Great Hall” highlights the improvements made to ray-traced reflections. DLSS 4’s transformer model produces clearer, more realistic reflections due to superior denoising (ray reconstruction). This results in diffuse reflections around the specular ones, and the surface detail is retained to a higher degree.

The native render fails to denoise the reflections to the same degree as the transformer upscaling model with ray-reconstruction. Regardless, the older CNN model produces similar quality as native, while FSR and XeSS lag behind.

DLSS 3 vs. DLSS 4 Frame Generation: VRAM Usage

We benchmarked the three games on our GeForce RTX 4090 to check if there’s a performance gain using DLSS 4. There isn’t, but the VRAM usage does see a slight decrease using the updated model. Cyberpunk 2077 sees a drop of over 300 MB at 4K “Max” using DLAA and frame generation. Enabling upscaling produces reductions of 100-150 MB in graphics memory consumption.

Alan Wake 2 exhibits similar results. The VRAM usage is reduced by 100-150 MB when pairing DLSS 4 upscaling with frame generation.

Interestingly, Hogwarts Legacy shows the largest improvements in graphics memory consumption. The VRAM usage drops from 14.74 GB to 14.22 GB (>500 MB) at 4K “Max” using DLSS at performance mode with frame generation.

Areej Syed

Processors, PC gaming, and the past. I have been writing about computer hardware for over seven years with more than 5000 published articles. Started off during engineering college and haven't stopped since. Find me at HardwareTimes and PC Opset.
Back to top button