Nvidia "confirms" DLSS 5 relies on 2D frame data as testing reveals hallucinations

Daniel Sims

Posts: 2,460   +74
Staff
Connecting the dots: Early reactions to Nvidia's DLSS 5 were swift and skeptical, with some observers likening the technology to an Instagram-style filter applied over gameplay footage. Nvidia CEO Jensen Huang refuted the allegations, but subsequent clarifications have helped outline how the system actually works – and where it can fall short.

According to Nvidia's press materials, DLSS 5 enhances lighting and material interactions without modifying underlying geometry. The company later reiterated this point to HotHardware, while TechSpot also highlighted commentary from Ryan Shrout arguing the tech is not simply a post-processing filter. However, in email exchanges with YouTuber Daniel Owen, Nvidia spokesperson Jacob Freeman clarified that DLSS 5 does not rely on an explicit understanding of 3D scene data for lighting or characters.

That explanation aligns with both Nvidia's official description and Shrout's analysis: DLSS 5 infers lighting and material behavior by analyzing individual frames rather than reconstructing full scene geometry. Digital Foundry has since confirmed that this approach can introduce errors, particularly when objects become occluded. In some cases, shadows can flicker or disappear when characters move across them.

Owen's testing highlights additional edge cases. While early criticism centered on a widely shared image from Resident Evil Requiem, another example from Starfield presents more clear-cut anomalies. Some viewers interpreted these issues as the system altering facial features, but closer inspection suggests most differences stem from changes in lighting, shadowing, and simulated skin translucency.

The now infamous Resident Evil comparison itself is somewhat misleading. Nvidia's promotional slider juxtaposes two different frames, where characters have shifted or left the scene entirely. Interestingly, the video thumbnail uses a separate frame pairing that, when aligned, shows DLSS 5 largely preserving the original character model.

The Starfield example is harder to dismiss (above). In a shot of two characters wearing spacesuits, the man on the left exhibits visible inconsistencies – his hair appears altered, and a shadow near his right nostril expands in a way that resembles a change in facial structure. When pressed on these issues, Freeman acknowledged that DLSS 5 remains an early preview. The system is not designed to modify scene details, but it can do so unintentionally.

What remains less clear is how the model was trained. Critics, including Owen, have speculated that DLSS 5 may rely on broad internet-sourced data, similar to large language models like GPT. Nvidia, for its part, has only described the system as a "unified model," without clarifying whether it builds on earlier DLSS architectures.

Developer involvement is another open question. Nvidia's announcement includes endorsements from Bethesda, Ubisoft, and Capcom, and like prior DLSS versions, the technology likely requires some level of manual integration. However, anonymous developers from Capcom and Ubisoft told Insider Gaming they were unaware of DLSS 5 prior to its reveal.

Even with adjustable intensity and exclusion controls, the risk of visual inconsistencies and potential for hallucinations remains. If those artifacts persist, they could complicate how developers balance DLSS 5's visual gains against the integrity of a game's original artistic direction.

Permalink to story:

 
You don't have to guess at what it is. I'm telling you that it's an upscaler paired with a diffusion model. It looks exactly like a diffusion model and there is no other technology in the world which it could be confused with.

I feel like I'm watching someone point at grass and ask how the landscapers make it green. It's grass!
 
Many of those people complaining about how it looks, would they have said anything if that was the original look of the characters?....I don't think so.

People complaining it's just because NVIDIA have shown a better way to do facial details than the developers themselves.
 
Last edited:
Many of those people complaining about how it looks, would they have said anything if that was the original look of the characters?....I don't think so.

People complaining it's just because NVIDIA have shown a better way to do facial expressions than the developers themselves.
There was no change to facial expressions though. There is only facial enhancements visually. Remember the artists can't request for no makeup if the dlss 5 adds it automatically. The choice for developers or artists is to use dlss 5 on in your game or off there is no middleground as mentioned in the q and a session. What kind of choice is that?
Also about 80% of polls show a rejection of the dlss 5 direction so not a handful of people but the majority of the market!
 
Last edited:
There was no change to facial expressions though. Thrre is only facial enhancements visually. Remember the artists can't request for no makeup if the dlss 5 adds it automatically. The choice for developers or artists is to use dlss 5 on in your game or off there is no middleground as mentioned in the q and a session. What kind of choice is that?
Also about 80% of polls show a rejection of the dlss 5 direction so not a handful of people but the majority of the market!
My bad, I meant to say facial details....still you can't deny the level of detail on their faces especially on Starfield characters are like day and night difference....in a great way.
 
Many of those people complaining about how it looks, would they have said anything if that was the original look of the characters?....I don't think so.

People complaining it's just because NVIDIA have shown a better way to do facial details than the developers themselves.
Better? The whole thing I saw looks like it's been dumped thru a Snapchat selfie filter from 2019, then smothered in Vaseline to make it all shiny. It also needs two 5090's to run, so....great. I also get to spend more money buying Nvidia's overpriced GPU's?

Nvidia's slogan used to be "The way it's meant to be played". The only ones being played is anyone buying into this AI crap they keep trying to justify.
 
Many of those people complaining about how it looks, would they have said anything if that was the original look of the characters?....I don't think so.

People complaining it's just because NVIDIA have shown a better way to do facial details than the developers themselves.
That's the problem, Nvidia is changing the look of characters and art design without the original developers or characters artist's permission. It's like taking a photo of a painting and throwing an AI filter over it and saying it's "better".

The only way any of these images are better is if someone prefers the modern direction of images and movies having AI filters and the effect of shoving a ring light in someones face.
Shadows and lighting is completely off, especially in the Starfield image, and the Fifa example is hilariously bad with the effects looking terrible, the blur and smearing is obvious as well.
 
That's the problem, Nvidia is changing the look of characters and art design without the original developers or characters artist's permission. It's like taking a photo of a painting and throwing an AI filter over it and saying it's "better".

The only way any of these images are better is if someone prefers the modern direction of images and movies having AI filters and the effect of shoving a ring light in someones face.
Shadows and lighting is completely off, especially in the Starfield image, and the Fifa example is hilariously bad with the effects looking terrible, the blur and smearing is obvious as well.
One thing for sure is that I see something that you are not seeing and viceversa. I absolutely love the level of detail especially on Starfield that goes from stretch out, no wrinkles on characters faces to something more on the lines on an actual human would have on their faces.
 
Better? The whole thing I saw looks like it's been dumped thru a Snapchat selfie filter from 2019, then smothered in Vaseline to make it all shiny. It also needs two 5090's to run, so....great. I also get to spend more money buying Nvidia's overpriced GPU's?

Nvidia's slogan used to be "The way it's meant to be played". The only ones being played is anyone buying into this AI crap they keep trying to justify.
Please try to be objective, we are talking about the technology features not the monetary value that would require to run it... if you had a ton of money the monetary cost will be the absolute last concern. With that said I do say that while it looks better, it is an insult as it changes the character actual looks way beyond than adding details.
 
Last edited:
Please rey to be objective, we are talking about the technology features not the monetary value that would require to run it... if you had a ton of money the monetary cost will be the absolute last concern. With that said I do say that while it looks better, it is an insult as it changes the character actual looks way beyond than adding details.
You work for Nvidia?
 
Many of those people complaining about how it looks, would they have said anything if that was the original look of the characters?....I don't think so.

People complaining it's just because NVIDIA have shown a better way to do facial details than the developers themselves.
Yeah, but it's not the original look. That's precisely the problem, genius, you hit the nail on the head.
 
I did admit that it's an insult to developers as it changes the characters' faces way beyond than just adding details, especially on Resident Evil....now it's your turn to admit that the actual details are great.

No. The "actual details" are a lie, since that isn't what the artists, modelers, and texturers MADE. It doesn't matter if the details look "great" since they are a lie. The hair wasn't supposed to be there. Dude had SHAVED sides of his head, not a fade.

It's literally just a stupid AI diffuser replacing the original rendering. It is NOT a rendering, it is not artwork, it's a forgery of the artwork. Nothing you say can refute this. You're really bad at all of this and being an Nvidia apologist here isn't going to help your case - since you don't have one.
 
My bad, I meant to say facial details....still you can't deny the level of detail on their faces especially on Starfield characters are like day and night difference....in a great way.
No worries, no one is denying the facial details are more enhanced. Some are saying and I agree that the difference is an overshoot of the whole environment or feel of the game. Also their are games with detailed faces without the need of 2 rtx 5090s like Calisto protocol and others. The gama levels were raised to make it look drastic again at the cost of brute forced approach using 2x 5090s. Initially let's recall they sold neural network as needing less vram.
Also let's recall what dlss stands for Deep Learning Super Sampling; dlss 5 isn't improving frame rates and not using upscaling only using a brute forced approach via 2x 5090s.
Visual quality of games are inevitably going to improve especially with 2 other players like Sony and Microsoft going all in on Neural network rendering. Nvidia just wants to corner the market with developers via quarter baked idea that is brute forced using 2x 5090!
Here is a reference of games using high detail faces before dlss 4 at native 4k. ( No 2x 5090 needed)1000056230.jpg1000056224.jpg1000056228.jpg1000056226.jpg
 
I did admit that it's an insult to developers as it changes the characters' faces way beyond than just adding details, especially on Resident Evil....now it's your turn to admit that the actual details are great.

Like I mentioned in an earlier post, you can no longer trust that photographs are accurate anymore...they are "enhanced" to look more pleasing. Obviously, you like the "enhancements" Nvida has done, whether through AI or algorithm, however, it is still , no longer the original art. The problem is, who is correct? And, is there a problem with all games being treated to the same modifications? At what point will all games start to "look alike" as this "improved" system takes hold?

Which brings me to how vociferously you argue that YOUR opinion is valid and the others live in the dark ages. Your opinion is neither right or wrong, but neither is anyone elses.
 
The choice for developers or artists is to use dlss 5 on in your game or off there is no middleground as mentioned in the q and a session. What kind of choice is that?
Not even remotely correct. Developers can adjust the level of enhancement applied by DLSS5, and even allow or disallow it for specific scenes or models within a scene.

Yeah, but it's not the original look. That's precisely the problem, genius, you hit the nail on the head.
You've missed the point entirely. If the developers are applying the effect, then the result is indeed the "original look", just as when a developer applies a texture, the "original look" of the underlying model is altered.

Strictly speaking, *any* form of interpolation, anti-aliasing, or upscaling is altering pixels in a manner not specified exactly by the developer. So? These are videogames, not signed Picassos.

You work for Nvidia?
The last refuge of the intellectually flailing disputant: accuse the other side of special interest.
 
Like I mentioned in an earlier post, you can no longer trust that photographs are accurate anymore...they are "enhanced" to look more pleasing. Obviously, you like the "enhancements" Nvida has done, whether through AI or algorithm, however, it is still , no longer the original art. The problem is, who is correct? And, is there a problem with all games being treated to the same modifications? At what point will all games start to "look alike" as this "improved" system takes hold?

Which brings me to how vociferously you argue that YOUR opinion is valid and the others live in the dark ages. Your opinion is neither right or wrong, but neither is anyone elses.
Can't argue your point, but this type of enhancements will live and die based on people preference to used them....just like the damn dog ears on phone filters.
 
Not even remotely correct. Developers can adjust the level of enhancement applied by DLSS5, and even allow or disallow it for specific scenes or models within a scene.


You've missed the point entirely. If the developers are applying the effect, then the result is indeed the "original look", just as when a developer applies a texture, the "original look" of the underlying model is altered.

Strictly speaking, *any* form of interpolation, anti-aliasing, or upscaling is altering pixels in a manner not specified exactly by the developer. So? These are videogames, not signed Picassos.


The last refuge of the intellectually flailing disputant: accuse the other side of special interest.
The cat is already out of the bag, anything remotely different from giving the artists control, some level of end user control via impact ( not just on and off), as well as efficiency will likely be rejected by the market imo. Who wants $10,000 beer goggles experience?
 
Last edited:
- There is not much room for artistic intent if the engine limits that. The AI enhances graphics and boosts it beyond hardware limitations
- Play first without judging, games are not chess boards.
- Little details, like make up? You are dealing with human characters, can happen. Unless dealing with characters that obviously have some physical and costume traits, like wonder women, for example that must 99% be preserved.
- Some studios don't focus on work of art type of developer intent, only want to give human like expression as lively as possible.
- Most games aren't work of art to begin with, they are titles, adventures and action stories. Slight shade or little change in detail won't change the whole narrative.
- You aren't playing photo sliders.
- Talented studios would know what to change in the tool. They play the game thousands of times before releasing it.
-They already stated that developers can choose what AI can address. But human faces usually gives the strongest impression. The way objects fit in graphical environment also plays it's part and are usually thousands of tiny details that Dev might not have time to deal with or even add. This tool might help them out.
 
Last edited:
Why did they call it DLSS? They could have called this tech whatever they wanted, why did their marketing team have to associate it with DLSS?
Because it uses neural networks built on Tensor Cores just like super sampling and frame generation. What this new thing is really is an amalgamation of those two technologies put together and likely running on the same Transformer Neural Network as DLSS 4+ does
 
Back