@Rich_Leadbetter I don't recall hearing anything specifc about DLSS 5's ability to maintain objective physical accuracy that may be present in the underlying rendering, nor discussions on whether a post-processing approach to enhancing the accuracy of lighting is a valid one, within the Q&A. If they were present and I misinterpreted or otherwise missed it, I apologise and if I didn't clearly convey those as the point of my comment then I apologise for that too.
Regardless of that, it's good to see you and your colleagues respond to the criticism well and openly learn from it, despite all the harassment it was mixed in with. It must not have been easy to keep a level head in that situation.
@Rich_Leadbetter What is your reasoning behind the claims that DLSS5 improves lighting or that it specifically is a leap forward in realistic rendering for environments?
For context, I know a modern game with a principled approach to physically based rendering takes the direction and authored material properties of a surface, alongside the colour, angle of approach, and strength of the light interacting with it, and uses that information in calculations designed to create outputs that match measured data of how real-world materials interact with light within as small a margin of error as plausible.
It seems DLSS5 can only take an image and alter it to visually match what is seen in a data-set of imagery that may not reflect anything seen within the game's world, let alone the specific lighting conditions. It can be seen outright removing shadows cast by the sun from a character's face in Starfield in favour of lighting that resembles photography that uses very specific lighting to highlight a subject and, while you've discussed issues with faces specifically, that seems to be a bit more than just an issue with the details of faces alone. Similar issues seem to be noticeable in environments too and it certainly isn't just tone-mapping. These are massive problems for the plausibility of lighting alone and that's before considering the fact that we can't ensure objective components of accurate lighting like energy conservation/preservation are present in DLSS 5's output and that it can't access information about material properties that could help it account for that.
Was any of the above considered during your reporting?
The possibility of machine learning having applications that improve the accuracy of modern lighting and rendering techniques is promising but the actual implementation of machine learning within DLSS5 seems rather useless in the pursuit of rendering that is accurate to the real-world.
Comments 2
Re: DLSS 5: Game-Changing Tech That Poses Big Questions For The Future Of Gaming
@Rich_Leadbetter I don't recall hearing anything specifc about DLSS 5's ability to maintain objective physical accuracy that may be present in the underlying rendering, nor discussions on whether a post-processing approach to enhancing the accuracy of lighting is a valid one, within the Q&A. If they were present and I misinterpreted or otherwise missed it, I apologise and if I didn't clearly convey those as the point of my comment then I apologise for that too.
Regardless of that, it's good to see you and your colleagues respond to the criticism well and openly learn from it, despite all the harassment it was mixed in with. It must not have been easy to keep a level head in that situation.
Re: DLSS 5: Game-Changing Tech That Poses Big Questions For The Future Of Gaming
@Rich_Leadbetter What is your reasoning behind the claims that DLSS5 improves lighting or that it specifically is a leap forward in realistic rendering for environments?
For context, I know a modern game with a principled approach to physically based rendering takes the direction and authored material properties of a surface, alongside the colour, angle of approach, and strength of the light interacting with it, and uses that information in calculations designed to create outputs that match measured data of how real-world materials interact with light within as small a margin of error as plausible.
It seems DLSS5 can only take an image and alter it to visually match what is seen in a data-set of imagery that may not reflect anything seen within the game's world, let alone the specific lighting conditions. It can be seen outright removing shadows cast by the sun from a character's face in Starfield in favour of lighting that resembles photography that uses very specific lighting to highlight a subject and, while you've discussed issues with faces specifically, that seems to be a bit more than just an issue with the details of faces alone. Similar issues seem to be noticeable in environments too and it certainly isn't just tone-mapping. These are massive problems for the plausibility of lighting alone and that's before considering the fact that we can't ensure objective components of accurate lighting like energy conservation/preservation are present in DLSS 5's output and that it can't access information about material properties that could help it account for that.
Was any of the above considered during your reporting?
The possibility of machine learning having applications that improve the accuracy of modern lighting and rendering techniques is promising but the actual implementation of machine learning within DLSS5 seems rather useless in the pursuit of rendering that is accurate to the real-world.