Gabriele Leone

Director of Content Tech @NVIDIA
DLSS 5 Developer.

[Gabriele Leone @ambertanuki]

@JoeWilsonEQ1 You posted an entire breakdown that was technically inaccurate, and your responses are coming across as disingenuous. You’re speaking confidently about something you don’t seem to fully understand, so I don’t think continuing this conversation is going to be productive.

[Gabriele Leone @ambertanuki]

@JayDook Well said :)

[Gabriele Leone @ambertanuki]

What you are describing is not an unavoidable DLSS 5 “look.”
Desaturated colors can be controlled with Tone Intensity, which can be set to zero to retain the source image’s color response.

Exaggerated wrinkles usually come from overly aggressive Structure Intensity, particularly in unsupported modded examples where values are pushed beyond the intended range or multiple passes are stacked.

Developers can choose different models, tune both intensities, and mask characters or individual regions. Most examples circulating online use the leaked DLL without that developer-led integration. They are not representative of how DLSS 5 has to look.

[Gabriele Leone @ambertanuki]

@tristanmeere This breakdown is almost entirely inaccurate from a technical standpoint, so I’m not sure it can really be described as “great.”

[Gabriele Leone @ambertanuki]

No, flickering is not “fine.” But you are attributing flickering from an unofficial Death Stranding 2 mod.
DLSS 5 is designed to remain closely aligned with the rendered input. If that input contains TAA instability, ghosting, unstable shading, aliasing, or other temporal artifacts, those problems are preserved.

DLSS 5 is an appearance-enhancement stage, not a system for repairing arbitrary defects produced earlier in the rendering pipeline. It also relies on correct motion vectors, temporal data, settings, and engine integration.

A leaked DLL injected into an unsupported game is not representative of a developer-tuned implementation.

We cannot determine the source of this particular flickering from the edited comparison alone. In my experience however, the overwhelmingly majority originate in the source renderer rather than being introduced by DLSS 5 itself. You would need the original frames and integration details before making that attribution.

[Gabriele Leone @ambertanuki]

At this point, comparing DLSS 5 to Clarity simply because both can increase local contrast is misleading. A Clarity filter remaps tonal relationships already present in the image.

It cannot recognize skin, foliage, glass, or metal and generate different material-specific responses, such as subsurface scattering on a character or light transmission through leaves.

But describing every synthesized lighting or material cue as a “false hallucination” is not an accurate technical characterization.

These cues are strongly conditioned on the rendered scene, not arbitrarily invented by an unconstrained image generator.

We also are not claiming that an untouched default output automatically represents the artist’s intention.

The output is art-directed by the game’s developers through model selection, Structure and Tone controls, semantic controls, and developer-authored masks.

At present, NBA 2K27 is the actual developer-tuned implementation you can judge.

The other games circulating online are unofficial experiments made by modders using the leaked DLL.

I was an artist in the games industry myself.
My team and the wider DLSS 5 dev-team include highly experienced industry artists, including former lead artists from Naughty Dog, Kojima Productions, id Software, and Rockstar Games, just to name a few.

Artists are not an afterthought here. They actively contribute to how the technology and its controls are developed.

[Gabriele Leone @ambertanuki]

@Brodizzleee @hutchinson Not sure what your screenshot is meant to demonstrate.
Yes, I just created my X account.
I am a developer for DLSS 5. Hope that helps.

[Gabriele Leone @ambertanuki]

DLSS 5 can increase local contrast, but as a consequence of generating the appearance of tighter contact shadows, ambient occlusion, reflections, subsurface scattering, and material-specific lighting.

A post-processing contrast or Clarity adjustment modifies existing pixel relationships across the image.

It does not understand where two objects make contact, where light should scatter through skin, or how different materials should respond toillumination.

“Hallucinating” is also too broad as a comment and does not reflect the reality of what DLSS 5 is doing.
It is not arbitrarily inventing new objects or geometry.

[Gabriele Leone @ambertanuki]

I agree that this particular result does not look the best, but you are attributing a modded configuration to an official implementation.

The exaggerated wrinkles, sheen, and facial contours are what happens when Structure Intensity is pushed beyond its intended range, sometimes to values such as 200%, or when multiple DLSS 5 passes are stacked.

That heavily amplifies high-frequency detail and is not representative of the settings normally chosen for a character’s face.

“We can only judge what we have” would make sense if this were a developer-native implementation. It is not. You are looking at an unofficial Death Stranding 2 DLSS 5 mod configured by a modder. Neither us nor Kojima implemented or tuned what you are judging here.

Developers have multiple models, Structure and Tone Intensity controls, automatic semantic masking, and engine-authored masks. They can reduce the effect on faces, apply different settings to characters and environments, or exclude specific regions entirely. This screenshot if anything demonstrates an excessively configured mod, not how DLSS 5 must look in a supported integration.

[Gabriele Leone @ambertanuki]

@DogParkSandbox That is true. In the case of DLSS 5 the better the input the better the output :)

[Gabriele Leone @ambertanuki]

Some effects can be baked or approximated, but they may not respond correctly as characters, lights, and cameras move.
DLSS 5 is not just adding contrast. It can generate the appearance of tighter contact shadows, ambient occlusion, subsurface scattering, reflections, and material-specific light response.
The limitation is not necessarily what the artists wanted to create. It is often what the renderer and available hardware can represent and calculate within the frame budget.

[Gabriele Leone @ambertanuki]

You are interpreting a lighting and tone change as an identity change. The facial proportions, silhouette, eyes, nose, mouth, and beard remain aligned. DLSS 5 simply giving the cheeks and nose a brighter lighting response. A person looking different under different lighting does not make them a different person. Developers can also control this directly. "Tone Intensity" adjusts broad lighting and color response, and setting it to zero retains the source image’s tone and colors. It can also be finely modulated or masked, so this particular setting is not an unavoidable result.

[Gabriele Leone @ambertanuki]

A Clarity filter only adjusts contrast around existing details. It cannot generate context-dependent appearance cues such as contact shadows, ambient illumination, reflections, material-specific light response, or light scattering through skin, hair, and foliage.

So no, you cannot replicate what DLSS 5 is doing in Lightroom.

[Gabriele Leone @ambertanuki]

You’re confusing training-time supervision with runtime inputs. DLSS 5 does not require developers to supply separate normal, albedo, and lighting passes to the deployed model. Those renderer-derived attributes are used during training to teach the model which parts of the scene should remain consistent; they are not runtime inputs that disappear without a “custom integration.”

At runtime, DLSS 5 uses the rendered RGB frame, engine motion vectors, carried temporal state, developer settings, and applicable control masks.

So it does not fall back to a different “simple img2img” mode when separate G-buffer passes are unavailable.
In the broadest sense, it transforms an input image into an output image. But calling it a simple img2img diffusion model omits the important parts: it is a specialized, deterministic, one-step pixel-space model with temporal conditioning and renderer-specific training not a general-purpose generator freely repainting an isolated screenshot.

The poor results from unsupported leaked-DLL experiments in older games do not prove your explanation. They can result from out-of-distribution inputs, incompatible integration data, missing developer tuning and masks, extreme settings, repeated passes, or other unsupported modifications.
Attributing them specifically to missing normal and lighting buffers is speculation based on an incorrect description of the runtime pipeline.

[Gabriele Leone @ambertanuki]

DLSS 5 is an artist-directed developer tool, not a universal filter automatically imposed on every game.

Developers select from different appearance models, adjust structure and tone intensity, control where enhancement is applied with per-pixel and semantic masks, and can reduce or disable it on specific regions wherever preserving the original look is more important.

None of that authoring process is represented by experiments using a leaked DLL. Modders are free to explore, but maximum-strength settings, unsupported configurations, and stacking many DLSS 5 passes are not representative of a properly integrated, developer-tuned implementation.

Likewise, a normal supported DLSS 5 pass cannot simply be prompted to redesign Cloud as 'Scarlett Johansson' or replace a stylized character with an unrelated realistic person.

It is strongly conditioned on the rendered frame and designed to preserve character identity, silhouettes, composition, and scene structure. Clips showing character replacement should be examined for fabricated content, additional AI tools, or extreme multi-pass experimentation. Ten stacked passes are *not* an example of what one intended DLSS 5 pass produces.

DLSS 5 currently targets photorealistic artwork and is not intended for strongly cel-shaded, cartoon, or voxel-based games. When applied to a stylized game changes may be limited mainly to local lighting such as ambient occlusion.

[Gabriele Leone @ambertanuki]

@3kliksphilip Was great to meet you at Gamescom and great video!

[Gabriele Leone @ambertanuki]

@opinali I enjoyed your thread. That said, DLSS 5 never actually changed the geometry, and that was true back in March as well. The current models are definitely better than they were then, though, and they’ll only continue to improve.

[Gabriele Leone @ambertanuki]

You have to keep in mind that the implementations you’re seeing online right now are being done by modders, not by the actual game developers. The developers have full control over which DLSS 5 model is used, where and how it’s applied, and all of the associated settings, including how it handles faces. There are specific controls for that.
I love the modding community, and I think it’s one of the coolest aspects of PC gaming, but it’s important to understand that these current implementations you find online don’t reflect how the original developers of those games would use DLSS 5 if they implemented it themselves.

[Gabriele Leone @ambertanuki]

@PellyNV @ctnzr 🚀🚀🚀

[Gabriele Leone @ambertanuki]

@edliu1105 Totally random! 😏

[Gabriele Leone @ambertanuki]

So excited to finally see this video out in the world! DLSS 5 is a huge leap forward for real-time rendering.

[NVIDIA GeForce @NVIDIAGeForce]

DLSS 5 introduces 3D-Guided Neural Rendering, which uses AI to infuse game scenes with lifelike lighting and materials. Developers tune the final output for each game to bring their artistic vision to life.