DLSS 5 in Starfield has become the rare case where a developer’s enthusiasm and player reality diverge sharply. Todd Howard, Bethesda’s executive, called the implementation “amazing,” yet the internet is calling it uncanny AI slop—particularly the AI-generated NPC faces that look wrong in ways players struggle to articulate but instantly recognize. This disconnect matters because it signals a fundamental disagreement about what constitutes acceptable AI-assisted graphics, and whether performance gains justify visual compromises.
TL;DR: Todd Howard praised DLSS 5 in Xbox Starfield as amazing, but widespread player backlash criticizes its uncanny NPC visuals as AI slop. The divide highlights tension between developer optimization goals and player perception of upscaling quality, with native DLSS now available on PC alongside mods that let older GPUs access the technology.
The Todd Howard endorsement versus player reality
When Todd Howard calls DLSS 5 in Starfield “amazing,” he is speaking from an optimization perspective—the technology delivers performance gains that allow the game to run smoothly on Xbox hardware. That is objectively true. But “amazing” and “acceptable to players” are not synonymous. The backlash centers on NPC faces, where DLSS 5’s frame generation appears to hallucinate details that look subtly wrong: eye placement shifts between frames, skin texture warps unnaturally, and expressions flicker in ways that trigger uncanny valley responses. These are not bugs in the traditional sense; they are artifacts of how AI frame generation reconstructs human faces from limited source data.
The core issue is that Bethesda optimized for performance metrics—frame rates, GPU load, visual fidelity at distance—without accounting for how players perceive close-up NPC interactions. In dialogue scenes, where the camera lingers on character faces, DLSS 5’s weaknesses become unavoidable. Todd Howard’s “amazing” reflects developer priorities; player criticism reflects user experience priorities. Neither is wrong, but they are measuring different things.
Why DLSS 5 in Starfield struggles with NPC rendering
Frame generation works by analyzing two rendered frames and inferring what the frames between them should look like. For static geometry—walls, terrain, distant objects—this inference is reliable. For complex, high-frequency details like human faces, it fails more often. Eyes are particularly vulnerable because they are small, expressive, and move in ways that require precise temporal consistency. When DLSS 5 reconstructs an eye position or blink timing from incomplete data, the result feels off in a way players cannot always articulate but immediately sense.
Starfield’s NPC crowd system compounds this problem. The game renders many characters simultaneously, often at distance or in peripheral vision, where DLSS 5 has more latitude to approximate. But when players approach NPCs for dialogue, the upscaler is forced to maintain quality on faces it was never designed to prioritize. The recent Starfield patch included specific visual fixes for “crowd eyes,” suggesting Bethesda recognized this weakness. Yet the patch did not eliminate the uncanny effect—it merely reduced its frequency.
AMD FSR2, which shipped with Starfield at launch, uses a different reconstruction method that produces fewer temporal artifacts on faces but delivers lower overall image quality and performance. DLSS 5 trades face consistency for frame rate—a calculation that Todd Howard found worthwhile, and that some players accept, but that others reject entirely.
Native DLSS support and the modding history
Starfield originally launched without native NVIDIA DLSS support, forcing PC players to choose between AMD FSR2 or using mods to retrofit DLSS functionality. The modding community responded with tools like FSR2Streamline and UpscalerBasePlugin, which intercepted FSR2 calls and replaced them with NVIDIA DLL files, effectively enabling DLSS without official support. This workaround required manual file placement and in-game configuration but delivered superior performance and image quality compared to native FSR2.
Bethesda has since released an official patch adding native DLSS support to PC, including DLSS Super Resolution, DLAA, NVIDIA Reflex, and Frame Generation for compatible NVIDIA GPUs. This legitimizes what modders were already doing and removes the friction of manual installation. However, it does not retroactively improve how DLSS 5 handles NPC faces—it simply makes the technology easier to access. Players who disliked modded DLSS can now dislike official DLSS with less effort.
The performance-versus-perception trade-off
Frame generation is fundamentally a performance technology. It allows hardware to maintain higher frame rates by rendering fewer frames and inferring the rest. On a technical level, this is a breakthrough—DLSS 5 delivers marked performance improvements over FSR2.2. But performance is not the only metric that matters to players. Temporal stability, facial consistency, and absence of visible artifacts matter too, especially in a narrative-driven RPG where players spend significant time looking at character faces.
Todd Howard’s “amazing” assessment likely reflects the fact that DLSS 5 solved Bethesda’s optimization problem. Xbox Starfield can now hit its target frame rate without dropping visual settings. That is a win for the studio. But players are evaluating the same technology by a different rubric: Does it look right? Does it feel immersive? Does it break the fourth wall by making AI artifacts visible? By that measure, DLSS 5 in Starfield is not amazing—it is a compromise that works for performance but fails for perception.
What this divide reveals about AI upscaling’s future
The DLSS 5 backlash is not really about DLSS 5. It is about the gap between what AI reconstruction can do and what players will tolerate. As upscaling technology becomes more aggressive—rendering fewer pixels and inferring more—the artifacts will become harder to hide. Faces are the canary in the coal mine because humans are exquisitely sensitive to facial irregularities. If frame generation can fool the eye on terrain and architecture but fails on faces, then there is a category of content where AI inference is not yet ready for player scrutiny.
This does not mean DLSS 5 is bad technology. It means DLSS 5 is not yet appropriate for all use cases. A competitive shooter where players see character models at distance or in motion might never notice the artifacts that dialogue scenes expose. A strategy game with minimal close-up character work would be unaffected. But Starfield, which is built around NPC interaction and dialogue, is precisely the wrong showcase for frame generation’s current limitations. Todd Howard called it amazing anyway, and that mismatch is the real story.
Can DLSS 5 be fixed for Starfield?
Bethesda could reduce NPC-specific artifacts by adjusting which frames DLSS 5 is allowed to generate. Disabling frame generation during dialogue scenes, or using a lower aggression setting for close-up faces, would preserve performance on the rest of the game while protecting the parts players scrutinize most. The recent patch that improved crowd eye rendering suggests Bethesda is aware of the problem and willing to iterate. Whether further patches will address the deeper issue of temporal face consistency remains unclear.
Alternatively, NVIDIA could improve DLSS 5’s face-specific reconstruction in future versions. Machine learning models can be trained to recognize faces and apply different inference rules to facial regions than to background geometry. This is technically possible but would require additional development and GPU overhead—costs that might outweigh the benefit for a single game.
Is DLSS 5 in Starfield worth enabling on PC?
If you have an RTX 40-series GPU and prioritize frame rate over temporal stability, yes. DLSS 5 delivers measurable performance gains over FSR2. If you spend most of your time in dialogue scenes and are sensitive to facial artifacts, no. Disable frame generation and use DLSS Super Resolution instead, which provides upscaling without temporal inference. If you are on an older NVIDIA GPU without Frame Generation support, mods enable DLSS upscaling without the AI artifacts.
Why did Todd Howard call DLSS 5 amazing?
Because it solved Bethesda’s performance optimization problem on Xbox hardware. Frame generation allowed the studio to maintain visual fidelity and frame rate simultaneously—a trade-off that works from a technical standpoint. Howard was not evaluating DLSS 5 from a player experience perspective; he was evaluating it from a developer perspective. The disconnect between those two viewpoints is the real story.
The DLSS 5 backlash in Starfield is a preview of a larger tension that will define AI upscaling’s next chapter. As the technology becomes more aggressive and more visible, developer optimization goals and player perception will collide more often. Todd Howard’s “amazing” is technically correct and player-experience wrong. That gap will not close until AI reconstruction can fool the human eye on faces as reliably as it does on walls and sky.
Where to Buy
Edited by the All Things Geek team.
Source: Windows Central
