Nvidia continues to take artificial intelligence to the next level with DLSS Ray Reconstruction. With her, she tries to improve ray layout in real time. Introduced with the RTX 50 series, this new evolution of DLSS replaces the traditional noise filters with an advanced AI model, generating more clear, detailed images and with a more realistic lighting, as we explain here …
I also recommend reading our article about the best graphics cards on the market
What is DLSS Ray Reconstruction?
DLSS (Deep Learning Super Sampling) Ray Reconstruction is a NVIDIA technology which uses artificial intelligence to improve the quality of images generated by Ray Tracing. It is part of the latest generation of DLSS 3.5, and also inherited by the new DLSS 4. A way to reduce noise in lightning layout in real time, as you know.
The purpose of DLSS Ray Reconstruction is to replace traditional den herishing filters, which eliminate noise but can degrade details, with an Deep Learning -based approach that improves the sharpness and global lighting of the scene.
For this to be possible, models of neuronal networks with a large number of rendering images and noise noise data should be trained so that it can differentiate between them, and thus generate an algorithm capable of intelligently predict how the final image would be like noise and thus apply to the generated frames.
Dlss ray reconstruction not only cleans the noise, but rebuild details based on global information From the scene. This allows more precise and less dependent images of the number of Ray Tracing samples, as I explained in another of our articles when we talk about Transformer Model …
I recommend reading our guide about the best gaming configuration
What is CCN?
This reconstruction or improvement of images can be made in several ways, one of them is the CNN and another by transforming model. To understand this, you must first say that Convolutional Neural Networks (CNNS) It is a deep neuronal network type designed to process images and detect spatial patterns. They have been widely used in deniew and super-resolution techniques in graphics.
For this technique, convolutionary layers are used to extract characteristics such as edges, textures and structures. Then the necessary changes to the image are applied and the noise is eliminated at the same time that the loss of detail is avoided as in more traditional techniques such as the so -called Denoisming.
What is Transformer Model?
The Transformer models They are a neuronal network architecture based on care mechanisms, originally designed for natural language processing, but now applied in graphics. In the context of DLSS Ray Reconstruction, the Transformers analyze the image in a more contextual and global way than a CNN. That is, it is smarter and with better results.
In this case, each pixel evaluates in Relaciñon with the whole scene, and not only with the adjacent pixels as in CNN. This provides a more intelligent reconstruction without loss of reflexes, shadows, or complex lights, in addition to reducing the dreaded artifacts in moving images, generating problems such as inconsistent animations or ghosting.
Advantages and disadvantages
https://www.youtube.com/watch?v=8ycy1ddgrfa
Now you will wonder what Advantages or differences There is between the Ray Reconstruction of NVIDIA based on CNN and the one based on Transformer Model, well, as a summary, to say that:
- Details: The CNN filters the noise and softens edges and textures, which makes losing some details and precision of the image, while the Transformer Model is more precise and clear. In addition, reflexes and lighting are more complex and without losses in the second case.
- Temporary coherence: CNN can generate ghosting in moving images, while stability is greater in Transformer Model.
- Ability: While CNN is more local in terms of the patterns that modifies, Transformer Model is more global for the entire full frame.
- Efficiency: CNN does not have a large impact on performance, while Transformer Model needs a somewhat higher calculation power.
Don't forget to leave your comments below …
