My Thoughts on DLSS 5.

Comments · 2.5K

• TR

  @TryToKevin · 3 weeks ago (edited)

  They really did it, they solved photorealism so hard that yearly sports games have become even more redundant.<br><br>Edit: This is indeed sarcasm.

  1.5K

• GM

  @gmb893 · 3 weeks ago (edited)

  <a href="https://www.youtube.com/watch?v=mnDvewMMzWM&amp;t=303">5:03</a> &quot;I turn it off and it looks like a game several console generations ago&quot; while playing GTA 5 which came out on the PS3 13 years ago

  1.3K

• IM

  @Im_TheSlime · 3 weeks ago (edited)

  strong opening<br><br>EDIT: i rarely use the position of top comment to state something but can everyone here stop being so negative and also stop trying to make shit up that sticks to philip?<br><br>the guy is literally the least problematic youtuber of all time and you got way too much spare time if you troll recreationally

  1.4K

• CA

  @cannotkazi · 3 weeks ago

  Once upon a time &quot;computer, enhance!&quot; was just a tired old movie trope

  265

• MA

  @Maldred_the_kind · 3 weeks ago

  DLSS5 Looks just like the moment before a migraine attack

  10

• MA

  @markackerman9485 · 3 weeks ago

  Its correct to say DLSS 5 isn&apos;t told by the engine what is and isn&apos;t a face, but assuming the training data was trained on a lot faces, then it essentially is recognizing faces and altering them more than world geometry.

  562

• SL

  @Slimek · 3 weeks ago (edited)

  I believe the model is tuned specifically for exaggerating face detail since the dll comes from NBA 2K27, and sports game fans are particularly obsessed with realistic faces (notably the whole &apos;cyberface&apos; genre of mods)<br><br>Edit: It&apos;s crazy that I have to clarify this but I am not attacking/disagreeing with Phil or anyone, I&apos;m simply sharing my point of view in this topic

  832

• VE

  @Venomhide · 3 weeks ago (edited)

  I feel the whole issues with faces and DLSS is that the lighting feels off on the face vs the environment. It feels like the DLSS lighting is &quot;White&quot;, faces get a white studio light shine, same for the environment. I doesn&apos;t feel right. The lighting doesn&apos;t seem to match things, like the faces having too much sheen (reflecting too much, or not enough light penetration into the skin). Several points even in the vid, the light intensity seems to be too harsh on faces, when the environment is overcast. I don&apos;t know what exactly the issue is, but it&apos;s the lighting mismatch to the environment that&apos;s triggering my uncanny DLSS&apos;essness.<br><br>Edit: To answer all the comments. The purpose of the comment is to point out why DLSS feels off (Faces especially), and its likely causes. Yes the tech is improving, yes it&apos;s technically jerry-rigged onto GTA/other games from what it was originally intended for. Identifying the reason WHY it feels off is just as important a discussion. We&apos;re trying to look past the elephant in the room of &quot;tech not made for game/not trained for this&quot;. <br>In order for true DLSS to work, especially in open world games like GTA, it&apos;ll need to be trained on all lighting situations, from natural world to office lighting to night clubs and candlelight, in order for it to not throw you off the second you change environments. How they&apos;ll do that and how well that&apos;ll work is another discussion.

  561

• BL

  @Blaze6108 · 3 weeks ago (edited)

  At <a href="https://www.youtube.com/watch?v=mnDvewMMzWM&amp;t=160">2:40</a> it&apos;s worth noting that these types of models <i>do</i> generate a new image, it&apos;s just very similar to the original. But if we&apos;re speaking technically as in this case, all models based on generative technology need to produce a new image, because their output is in the form of an image, not of a set of changes that can be applied to existing content like e.g. a classical filter would.

  12

• II

  @IIStaffyII · 3 weeks ago (edited)

  I think its worth pointing out its not only that we humans are good at detecting changed applied to human faces. It also that the models likely have disproportionate amount of data on human faces. Image models can overfit on a particular thing, which usually causes that particular thing to look like someone has used a HDR filter or played with the image contrast. Which makes it sorta useful to have the masking option, so one can have it less applied the areas that has received overfitting. Atleast it looks like overfitting to me. I&apos;m no expert, but done my fair share of image training back when SDXL was the most advanced thing around.

  68

• MR

  @mrcraggle · 3 weeks ago

  As someone with some experience in studio lighting, this has studio lighting look all over it. Of course, just modding it into any game that isn&apos;t fully supported isn&apos;t at you&apos;re not going to get the best experience but it also doesn&apos;t help that NVIDIA themselves in the techs first showcase gave yassified looks to certain people and again, the studio lighting issue remains there even in dark scenes, cloudy scenes, overcast scenes, everything seems to be lit in a certain way that looks like it has almost hero lighting.

  187

• AB

  @abdo.reda.m · 3 weeks ago (edited)

  With all due respect, I think there are some misinformation in this video. Increasing sub surface scattering? It&apos;s not doing anything to the underlying materials, it&apos;s literally changing pixels colors depending some sort of ai model. Not recognizing faces? I am pretty sure the whole image is fed into this model, while it operates per pixel, that has nothing to do with image convolutions or the input is to the model. I am just saying what I understand about how these things traditionally work.

  18