Final Fantasy XV DLSS versus TAA IQ and Performance Analysis


Final Fantasy XV DLSS versus TAA and Image Quality and Performance Analysis

This Final Fantasy XV DLSS patch, image quality, and performance analysis is a follow-up to BTR’s Game and Performance Review.  We found that the Highest settings look incredible but unfortunately they make even the RTX 2080 Ti struggle to play well at 4K with TAA anti-aliasing.

This morning, NVIDIA released a new Game Ready 417.35 driver and this evening, Square Enix also released a large 15.4GB content update with a multiplayer expansion and a brand-new quest which also enabled DLSS. So we will compare DLSS versus TAA performance and their respective image quality.   This beta DLSS update promises a substantial increase in DLSS 4K performance for Final Fantasy XV.

Deep Learning Super-Sampling (DLSS)

The RTX 2080 Ti Turing’s 114 TFLOP Tensor Cores are specialized execution units that perform tensor matrix operations that are the core compute function used in deep learning.  Deep Learning SuperSampling (DLSS) leverages a deep neural network to extract features of a rendered scene and it accurately combines details from multiple frames to construct a high-quality image. Turing GPUs use half of the samples for rendering and then use AI to fill in data to create a final image. The result is a less blurry image with similar quality to TAA but with higher performance.

Here is a comparison taken from the Final Fantasy XV demo which runs a timedemo at maximum settings using either DLSS or TAA.  The screenshots are similar but not identical.  The screenshot using DLSS is on the left and TAA is on the right.

Although the images are not identical, DLSS is on the left and TAA is on the right

In some places the DLSS image looks a bit sharper although it is hard to tell from this example.  The foliage looks much sharper but the license plate is a bit more blurry. The images are quite similar but the performance using DLSS is much higher than using TAA as the following chart shows from our original Turing launch review shows.

While TAA renders at the final target resolution and then combines frames, subtracting detail, DLSS allows faster rendering at a lower input sample count, and then infers a result that at its target resolution is of similar quality to the TAA result, but with roughly half the shading work.

The key to implementing DLSS involves its deep learning training process.  It actually learns how to produce the desired output based on large sample high-quality examples.  Thousands of reference images are collected and rendered with 64x supersampling (64xSS). 64xSS shades at 64 different offsets within the pixel, and then the images are combined to produce an “ideal” image. Matching raw input images are also captured and then rendered normally for comparison.

The DLSS network is then trained to match the 64xSS output frames, by going through each input using a complicated weighting process called “back propagation”. Eventually, DLSS is able to produce results that closely approximate the quality of 64xSS, while also avoiding blurring, disocclusion, and transparency that affect traditional post process anti-aliasing methods like TAA which is used for Final Fantasy XV.

DLSS 1X as used in Final Fantasy XV uses a render resolution of 1440p together with an upscaling algorithm to simulate 3840×2160, and we shall see if it is directly comparable to native 4K with TAA.  It is not yet possible to use DLSS at other than 3840×1440 unless you use DSR to upscale.

By using NVIDIA’s supercomputer to train their neural network which scans a game’s images, DLSS becomes a relatively easy integration for developers.  So far, developers have announced that 25 games will support DLSS so that gamers will be able to have DLSS anti-aliasing at a lower performance cost than by using TAA but with similar image quality.

The above comparison images are from the Final Fantasy XV demo.  Before we check the comparative image quality and performance between TAA and DLSS in the full game, let’s check out our test configuration on the next page.