| Otter said: Beyond A.I skeptisim though there is just genuine criticism over what its representing. Even 4 years ago this bottom image would look fake or photoshopped. |
I think this is the meat of your point, and I'd like to address this. I think this argument can be applied to many rendering technologies. For example, older games were designed with certain limitations in mind by their artists. When their textures are upgraded or let's say path-tracing is applied with something like RTX Remix, the atmosphere and style can (and almost certainly does) change drastically, and it's not always in line with say the original concept art (not that the original game is exactly reflective of concept art, either.) This "misrepresents" the original vision in the same way DLSS 5 seems to. And yes, people who are very purist about intentionality in their media consumption do boycott these sort of remasters. They represented a small part of a large spectrum. On the other extreme of the spectrum are people who love to mod their games, and this I think is a larger segment of PC gamers than the prior. Although I am mostly basing that on intuition more.
For new titles released after the technology itself, I don't think this same dichotomy exists. Nobody (or rather very few people) say(s) that Cyberpunk 2077 is misrepresented by path-tracing, even though the game looks drastically different with it on. That's because the game was developed with ray-tracing in mind, and it was a natural step to transition to path-tracing from there. It was a technology that was anticipated. This isn't the case with DLSS 5, and there will need to be time for artists and developers to catch up and experiment.
I also think people are extrapolating too much from what they know (or think they know) about pre-trained image and video AI models and applying it here, are some times confirming biases when analyzing an image, without considering all possibilities. The whole point of using a model trained on motion vectors towards deterministic targets is that the artists have more control over the results. On the other-hand vision and image models fail to achieve their goal when they are made too deterministic and that is a very real trade-off that results in very obvious "hallucinations." A model like DLSS 5 is neither trained like a diffusion image/video model, nor is it trained like an autoregressive next-token predictor. Its main goal is to match a ground-truth target using a lossy input of meaningful features. By its nature it is meant to be as deterministic as possible. There is no "creativity" element like with image and video generators. Yet I've been seeing people looking for "creative" aspects that I am not convinced are actually that. There are other explanations (like the model not accurately predicting the ground-truth; for example, mistaking a depth shadow as makeup), and those explanations make it likely that the ability to accurately meet a target ground truth improves over time.
Last edited by sc94597 - 17 hours ago
