The video dives into Nvidia’s DLSS 4, which replaces the older CNN model with a new Transformer model, significantly enhancing image quality in ray reconstruction, super resolution, and frame generation while remaining backward compatible with all RTX cards. Key improvements include better detail and realism in textures and skin rendering, along with reduced ghosting and smearing, although older RTX models may experience a drop in performance when using the new model.
The video discusses the launch of Nvidia’s DLSS 4, which introduces a new Transformer model to replace the older convolutional neural network (CNN) model used since 2020. This new model promises significant improvements in image quality for super resolution, ray reconstruction, and frame generation. Importantly, DLSS 4 is backward compatible with all RTX cards, allowing users to benefit from the upgrades regardless of their GPU generation. The video focuses on comparing the performance and quality of ray reconstruction between the old CNN model and the new Transformer model in various gaming scenarios, particularly in titles like Cyberpunk 2077 and Alan Wake 2.
One of the key areas of improvement highlighted is in ray reconstruction, which had previously faced issues with image quality, particularly in terms of smudging and ghosting. The video illustrates how the Transformer model addresses these issues by enhancing detail and reducing unwanted stylization, which often led to an overly sharp and artificial appearance in the CNN model. Viewers can see side-by-side comparisons showing how textures, such as tiled walls and wooden surfaces, exhibit more accurate detail and stability when using the Transformer model, resulting in a more natural and visually appealing output.
The video further explores how the new model handles skin rendering, noting that the CNN model often exaggerated skin details, giving characters a dry and aged appearance. In contrast, the Transformer model effectively simulates subsurface scattering, which softens the appearance of skin and enhances realism. This is demonstrated through comparisons of character models in both Cyberpunk 2077 and Alan Wake 2, where the Transformer model results in a more lifelike representation of skin, with improved color depth and reflection details.
Another significant improvement with the Transformer model is the reduction of ghosting and smearing, particularly in scenes with indirectly lit objects. The video showcases how characters and NPCs appear more stable and natural when moving, with less distortion and a more coherent representation of their features. This is especially noticeable in Cyberpunk 2077, where previously, the CNN model struggled with ghosting effects in certain lighting conditions, leading to an uncanny visual experience.
Finally, the video assesses the performance implications of the new Transformer model across different RTX GPU generations. While the model performs well on newer RTX 4000 series GPUs, older models like the RTX 2080 Ti and 3090 experience a more significant drop in frame rates when using the Transformer model for ray reconstruction. Despite these performance costs, the video concludes that the image quality improvements offered by the Transformer model are substantial, making it a worthwhile upgrade for users, especially when adjusting internal resolution settings to balance performance and quality. The presenter hints at a follow-up video focused on DLSS 4 super resolution, inviting viewers to stay tuned for more insights.