Hacker News new | ask | show | jobs
by Lord-Jobo 85 days ago
Watching a full video of A and B comparisons shows the massive weaknesses of the tech central to people’s complaints. (https://m.youtube.com/watch?v=YjvZMbsjEhY) skip to 3:46 and mute. You’ll see in the hogwarts examples, especially the last one with the dark background, that the lighting changes from something ‘logical’ to something nonsensical. In that, the character has multiple points of lighting on their face tied to NOTHING in frame and nothing in the scene. It goes from a dark image to a weird surreal tableau. In the assassins creed example you will see the complete deletion of key shadows, namely the ones cast from clouds moving overhead. This is purely a loss of detail, hard to argue the value for that. In oblivion you’ll see how it makes very bold guesses about textures for people’s faces and uhhh misses pretty bad with the high elf shown. In ALL examples it absolutely blasts contrast and crushes shadows to reduce dynamic range. For scary atmospheric games like resident evil this has a substantial impact on how they feel. In MOST examples, including the headline image of grace from RE9, it’s doing quite a bit more with the input data that simple upscaling. Her lips are a different size and color, her eyes have a different proportions. It’s not normal makeup and lighting changes(makeup alone is very controversial, her written character not wearing makeup literally makes way more sense she is a deeply traumatized antisocial introvert), the sizes of her features are literally different. It’s way more obvious in oblivion with the guy on the docks: his eyes are completely different textures. Looks like a distant relative of the original dude. The reason that last part is a problem even if you think the dlss5 version looks better, is how much impact you are ceding as a developer over to nvidias black box when you enable this. It’s a strong reduction in artistic clarity, from drawing board to people’s displays. And nvidia pushes REALLY hard for developers to use this stuff, at multiple levels. Game devs are artists who do not need even more pressure to sacrifice their art, that’s for sure. And finally, nvidia seems set on this, they are bundling this with their dlss version numbers. But there is nothing else in their dlss suite doing anything like this, in terms of impact. Bundling it is a deceptive tactic, many people like myself enjoy the upscaling even the frame gen sometimes, but would NOT tolerate this effect. If nothing changes, devs are going to be forced to choose between a) old worse upscaling/frame gen or b) newest best version of those, with the ugly “instagram filter” effect.

I am not an AI hater at all, I just like pretty games. And this is an absolutely wild misstep from nvidia. But a fixable one, if this model is tuned way down and retrained not to touch faces the way it does, I can see the appeal.