
On October 13, 2026, The Verge reported that Nvidia has dialled back the AI face generation in DLSS 5, the upscaling and neural rendering suite that will reach a commercially released game for the first time next month, through an update to Hogwarts Legacy. When DLSS 5 was shown earlier in 2026, it was quickly branded generative AI for games: a 15-year-old Hogwarts Legacy student was rendered visibly older, and Grace, the lead of Resident Evil Requiem, was given a different face. The rollback
The timeline is fairly clear. In January 2026 Nvidia unveiled DLSS 5 at CES as the starting point of neural rendering: beyond super resolution and frame generation, it uses a generative model to rebuild faces and materials, filling in skin texture, hair strands and lighting from a lower-resolution input. Two titles anchored the demonstration, Hogwarts Legacy and Resident Evil Requiem. The backlash focused on faces. In the demo, a 15-year-old Hogwarts Legacy student came out with visible nasolabial folds and loosened skin, while Grace, the protagonist of Resident Evil Requiem, was rendered as someone else entirely. The community labelled DLSS 5 generative AI for video games and argued that Nvidia had traded art direction for visual spectacle. For roughly ten months Nvidia did not publicly address the criticism. According to The Verge, the shipping version now reduces how far face
To understand the concession, look at how DLSS evolved. DLSS 2 in 2020 did super resolution only, rebuilding a high-resolution image from a lower internal resolution. DLSS 3 added frame generation in 2022. DLSS 4 switched to a transformer model and introduced multi frame generation in January 2025, and DLSS 4.5 added dynamic multi frame generation, up to 6x, in January 2026. That whole path was reconstruction, restoring information that already existed. DLSS 5 puts a generative model inside the render pipeline and lets AI create facial detail that was never captured or was downsampled away, which is a different contract. Meanwhile players have grown far more sensitive to AI-generated
Liu Gong's take: Nvidia is not abandoning DLSS 5, it is converting an always-on beautifier into a tiered, switchable feature. Within six months, face and material generation will most likely appear as a separate slider in the driver panel, and the default tier will be deliberately conservative.
From a technical standpoint, DLSS 5 takes a low-resolution input and lets a generative model complete it, which is a different contract from super resolution. In industrial inspection, where Liu Gong works, an algorithm is allowed to recover only what the sensor lost; if a model hallucinates a scratch that is not there, or paints over a real defect, the line stops. Games carry no safety risk, but the logic holds: the developer ships an art asset, and an invented nasolabial fold is added pixels that nobody authored. Lowering the strength is Nvidia admitting that generative output cannot override source data without limits.
On the business side, the arithmetic is simple. DLSS has pushed frame-rate multipliers from 2x to 4x to 6x, and each step buys less. What remains as a reason to buy a card is image quality. Hogwarts Legacy has sold more than 30 million copies, so any change to its rendering lands in tens of millions of living rooms, and one screenshot of an aged face can wipe out months of technical marketing. Frame generation artefacts such as ghosting are tolerated; a stranger's face is not.
For the industry, the concession sets a precedent: the strength of generative rendering has to yield to art direction. Liu Gong expects major engines to add AI-content flags and intensity metadata to their render pipelines over the next year, so studios can toggle the effect scene by scene, and vendors to market neural rendering on fidelity rather than beautification. Whoever standardises that control first will hold the argument in the next GPU cycle.
- DLSS frame-rate multipliers have grown by generation: 2x with DLSS 3 in 2022, 4x with DLSS 4 in January 2025, and up to 6x with DLSS 4.5 dynamic multi frame generation in January 2026, while face and material reconstruction adds no frames at all, only different pixels.
- Hogwarts Legacy launched in February 2023 and has sold more than 30 million copies, so the first shipping build of DLSS 5 reaches a user base in the tens of millions.
- Capcom's Resident Evil series has shipped over 170 million units, and Resident Evil Requiem was among the first DLSS 5 demonstrations.
- Roughly nine months separate the CES demo from the confirmed retuning, and Nvidia has not published the intensity tiers
Short term, over the next three months, how DLSS 5 looks in Hogwarts Legacy will set the adoption pace for everyone else. If comparison videos land well, more AAA titles follow in the first half of 2027; if the aged-face screenshots persist, most publishers will ship super resolution and frame generation only and hold generative faces for a later generation. Medium term, over a year, neural rendering is likely to become a tiered licence, toggled scene by scene and character by character, with AI-content labelling spreading from store pages into real-time disclosure.
For ordinary players, the near-term change is simple: faces in Hogwarts Legacy will look closer to the original release, with
Questions worth exploring:
- Whether generative rendering intensity becomes a standard driver control, and how it coexists with graphics presets
- Where responsibility sits when AI completion conflicts with authored art assets: studio, engine or GPU vendor
- Whether AI-content disclosure rules spread from store pages into real-time rendered frames
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from The Verge AI (published 2026-10-13 13:55:45). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.
