The internet has spent the last week losing its collective mind over DLSS 5. Screenshots of Cloud Strife looking like a SNL cast member, Skyrim characters suddenly transformed into Instagram models, and Halo 2’s Miranda Keyes getting a soap opera filter have fueled a predictable backlash: “AI slop is ruining our games.”
Then Daniel Vavra, the creative director of Kingdom Come: Deliverance 2, did something unexpected. He tested the leaked DLSS 5 library in his own game, posted comparison images, and told the internet it’s actually closer to his original artistic vision than the shipping game.
The responses ranged from “he’s lost it” to “he’s being paid by NVIDIA” to the usual chorus of “THIS IS NEURAL SLOP.” But here’s the thing: Vavra might be right, and the controversy reveals a fundamental misunderstanding about what DLSS 5 actually does under the hood.
The “Leak” That Started Everything
The DLSS 5 library wasn’t officially announced. Modders extracted the files from NBA 2K27(presumably the only game with early access to the tech), and within days, enthusiasts were forcing it into everything from Skyrim to PlayStation 2 emulators. The results were… inconsistent. Crank every setting to maximum, and you get the uncanny valley nightmares that dominate the discourse. But that’s like judging a car by flooring it in first gear.
Vavra’s approach was different. He loaded it into KCD2, a game known for its meticulous historical accuracy and photorealistic character models, and looked at what changed. The answer, according to his analysis, is that DLSS 5 fixes problems the game’s engine couldn’t solve, without altering the underlying geometry.
What DLSS 5 Actually Changes
Here’s where the technical details matter. Vavra’s comparison shows two images of the same character head model. Left side: the shipping game. Right side: DLSS 5 enabled. No geometry changes. Same mesh, same textures, same normal maps. What’s different?
- Skin detail: significantly more detailed skin and facial textures that were always in the assets but couldn’t be rendered properly
- Lighting: more realistic skin lighting that matches what the artists saw in ZBrush
- Ambient occlusion: enhanced, with shadows from hats, helmets, hoods, and small details like buckles and bags
- Hair rendering: more pronounced and darker shadows within hair strands
- Environment: minor improvements to shadows and vegetation textures
The key insight from Vavra’s technical explanation: the character head models were 3D scanned from real people using photogrammetry, then refined by artists. The final in-game result never matched the reference, not because the artists failed, but because the game engine’s lighting system couldn’t correctly display the shadow data, skin texture, and normal map details that were already in the assets.
As Vavra put it: “On the left side of the picture is a head spoiled by a very limited game engine lighting system that cannot correctly display shadows, skin texture, and normal map details. On the right is the SAME MODEL WITHOUT ANY GEOMETRY CHANGES, which looks exactly like a render from ZBrush, where the 3D artist modeled it, only with better lighting.”
The Problem Isn’t AI, It’s Game Engines
This is the part that gets lost in the “AI slop” discourse. Modern game engines are incredibly complex systems built to run on hardware that’s always a generation behind the artist’s intent. The artists at Warhorse Studios created photorealistic faces using ZBrush and photogrammetry, then spent months fighting the CryEngine to make those faces look right in-game.
Vavra’s frustration is palpable: “I often criticized the character artists’ work when certain faces in the game looked lacking in detail and plasticity, and they’d often show me that it looked fine on their computers and that the problem was with the lighting in the game.”
Translation: the artists were doing their jobs. The engine was failing them.
This isn’t unique to Warhorse. Every studio that pushes for photorealistic characters hits this wall. The geometry and texture data represent the artist’s intent, but the rasterization pipeline, with its limited shadow cascades, simplified lighting models, and approximations, can only approximate what the artist designed. DLSS 5’s neural network rendering essentially fills in the gaps, using its understanding of scene geometry and motion to produce the shadows and lighting that the engine should have generated but couldn’t.
The Frame Rate Elephant in the Room
Now for the part that gives Vavra’s endorsement some credibility. He didn’t pretend DLSS 5 is magic. His exact words: “The ‘minor’ issue is that it cuts the frame rate in half :)”
Let that sink in. A technology that’s supposed to be about performance, DLSS has always been marketed as a way to get higher frame rates through upscaling, is being used here for the opposite purpose. DLSS 5’s neural rendering pipeline is doing heavy lifting that the traditional rasterization pipeline can’t handle, and that costs compute. On high-end hardware, you might get away with it. On the Steam survey’s most common hardware, 1080p GPUs that are several generations old, you’re not getting playable frame rates.
This is a genuine tradeoff, and it’s the one legitimate criticism of DLSS 5 as a “fix” for engine limitations. Developers can’t ship a game that requires DLSS 5 to look correct unless they’re willing to tune the tech for their specific performance targets. Which is why Vavra’s assessment matters: he’s not saying “DLSS 5 is perfect”, he’s saying “DLSS 5 fixes rendering problems that have plagued us since launch.”
Why the “AI Slop” Narrative Fails
The backlash to DLSS 5 reveals something uncomfortable about how gamers discuss AI: most people have no idea how the technology works. As one developer-friendly voice noted, “At a core level those people don’t understand how DLSS or any of the more advanced technologies work. They’re not taking input image -> output image. They use a much deeper understanding of scene geometry and movement flow.”
This isn’t the same as generative AI inserting things that weren’t there. DLSS 5 takes the existing rendered frame plus motion vectors and scene data, then produces a higher-quality output. It’s not inventing new geometry or “fixing” faces into something different. It’s taking the data that was already there, the photogrammetry scans, the normal maps, the lighting calculations, and rendering them the way they were supposed to look.
The people calling it “AI slop” are often comparing the extreme settings (which do produce uncanny results) with the default settings (which are much more conservative). Vavra explicitly addressed this by noting that the most shocking DLSS 5 images online are using maximum settings across multiple parameters, creating visuals that “vary significantly from the original source’s tone and color schemes.”
The Bigger Question: Will Developers Use This as a Crutch?
The legitimate concern isn’t whether DLSS 5 makes games look better, Vavra’s comparison shows it clearly does. The real question is whether developers will start shipping broken games and expecting DLSS 5 to fix them in post. One commenter captured this anxiety: “It will be interesting if developers put less time and effort into creating the base art and key assets though, if they rely on the ‘we’ll fix it in post’ push magic, make everything look better button.”
There’s precedent for this fear. The history of game development is filled with technologies that were supposed to enhance artistry but ended up enabling laziness. Pre-built asset stores, middleware engines, and automated tools have all been accused of homogenizing game design. DLSS 5 could follow the same path.
But here’s the counterargument: the grunt work of modeling isn’t the artist’s vision. The artist’s vision exists in their head, and every tool that gets them closer to that vision faster is a win. As one thoughtful commenter put it: “If there was a button to immediately transfer an artist’s idea into a finished product, that would be true artistic freedom. Not spending thousands of hours in an engine.”
The question isn’t whether DLSS 5 makes games look better, it does. The question is whether developers will use it as a tool or a crutch. That’s a cultural question the industry will answer over the next few years.
What This Means for the Future of Game Development
For game developers and technical artists, Vavra’s endorsement is a signal that AI-assisted rendering is moving from “interesting experiment” to “production tool.” The implications extend beyond just visual fidelity:
-
Asset pipelines could become more efficient: If the rendering pipeline can finally display the detail that’s in photogrammetry scans, artists might not need to spend months manually painting detail that the engine should have rendered in the first place.
-
Engine limitations become less of a bottleneck: The gap between artist intent and what engines can render has always been one of game development’s dirty secrets. DLSS 5’s approach, using AI to fill in where the rasterization pipeline falls short, could finally close that gap.
-
Performance expectations will shift: As Vavra noted, DLSS 5 cuts frame rates in half. That’s a non-starter for many games, but it suggests a future where this kind of neural rendering is baked into the engine’s lighting model, not bolted on as an afterthought.
-
The “AI slop” discourse will mature: Eventually, people will realize that criticizing DLSS 5 for “changing” games is like criticizing HDR for making colors more accurate. It’s not changing the art, it’s finally rendering it correctly.
The cynics will say Vavra is just defending his game’s reputation. But he has nothing to gain from this, KCD2 has been out for months, and he’s already working on a Lord of the Rings RPG and another Kingdom Come project. If anything, he’s being too honest about the technology’s current limitations. As he told his followers: “You have no idea how terribly, and I mean really, really, annoyed I was that in KCD2 we couldn’t render shadows from hats, hoods, and helmets on characters’ heads, or even shadows from small details on character models. DLSS 5 does this, and it does it absolutely brilliantly.”
That’s not the perspective of someone defending a payday. That’s a developer who spent years fighting engine limitations, finally seeing his team’s original vision realized.
NVIDIA’s advancements in AI have been impressive across the board, from quantization techniques that beat FP8 to model compression strategies that make multi-billion parameter models practical. But DLSS 5 might be the first NVIDIA AI technology that genuinely solves a problem developers have been struggling with for decades: the gap between artistic intent and engine capability.
The “AI slop” narrative isn’t just wrong, it’s actively harmful. It conflates genuine technical problems (frame rate costs, inconsistent results at extreme settings) with aesthetic preferences being violated. When a respected developer like Vavra says “this is our original creative vision, and the model looks like static renders from ZBrush”, maybe we should listen instead of reaching for the nearest dismissive meme.
The real test will come when DLSS 5 ships officially and developers have to decide whether they’re using it to enhance their games or to excuse shipping broken ones. NVIDIA’s role in that decision, and whether they provide tools to help developers tune DLSS 5 for their specific performance targets, will determine whether this technology becomes a standard feature or another controversial footnote in gaming history.
