DLSS 5 explained, with Linus Tech Tips | The Vergecast

在最新一期的The Vergecast节目中，主持人David Amell欢迎了来自Linus Tech Tips的Linus Sebastian和David Gautier，共同讨论了英伟达（Nvidia）于9月3日发布的DLSS 5技术。三人深入探讨了为什么深度学习超采样的这一新迭代在游戏和开发者社区中引发了巨大争议。 历史上，DLSS旨在通过机器学习来放大低分辨率渲染，以填充缺失的像素，很像增强低分辨率照片。然而，Linus Sebastian解释说，DLSS 5超越了这一点，“它似乎不仅仅是，你知道的，添加了原本不存在的像素，它似乎在添加场景中原本不存在的元素，或者添加原本不存在的颜色。”英伟达明确表示，他们坚信“摩尔定律……已经彻底消亡”，而这种“神经渲染”才是未来。他们设想了一个世界，开发者只需简单地输入“桶”，一个照片般逼真的桶就能立即且廉价地生成，而无需传统地渲染所有像素。 然而，David Gautier在《Expedition 33》和一款《最终幻想》游戏中对DLSS 5的亲身体验，却呈现了另一幅画面。尽管他承认英伟达的愿景，但他指出当前的实现与“魔法桶”的理想相去甚远。Gautier将其比作“Instagram滤镜”，强调其最大的限制是需要每秒渲染60次，这限制了AI的“创造力”。他观察到，好的输入会产生更好的输出，但它仍然保留着一种“AI外观”，这可能会让那些厌恶AI生成内容的人产生“厌恶反应”。 讨论很快转向了对艺术意图的影响。Amell强调了Gautier的观察，即DLSS 5可能将一个风格化的动漫角色转变为“真人”，这可能会剥夺游戏独特的审美。当前的争议因测试人员使用的DLSS 5 DLL文件遭到泄露而变得复杂；目前只有《NBA 2K27》经过了英伟达的官方调优。Linus认为这项技术可能只对追求照片级真实感的游戏有利，而风格化游戏则会受到影响。Gautier还指出，DLSS 5有“吸走色彩”的倾向，类似于“电影中偷懒的色彩校正”。一些开发者，例如《Kingdom Come Deliverance 2》的创意总监，则为它辩护，声称它实现了他们原有的艺术愿景，只是当时缺乏资源去实现。 性能和成本是主要的争议点。DLSS 5带来了“灾难性的”性能损失，即使在高端的5090显卡上也是如此，据报道，早期的演示需要两块5090显卡。Linus质疑其价值主张，指出升级硬件（5080显卡价格为1500-1600美元）的“疯狂”成本。他认为英伟达的信息传递令人困惑，同时推广高帧率、高分辨率、光线追踪和AI增强，而当前硬件难以同时实现这些。Gautier认为英伟达有意让DLSS 5要求苛刻，以确保用户需要其性能提升，从而强化对其生态系统的依赖。这迫使习惯于在高端硬件上将设置调到最高的PC游戏玩家做出妥协。 两位嘉宾都承认英伟达正在对负面反应做出回应，例如有计划推出“DLSS 5的4000系列版本”。虽然他们期待改进，但Linus警告说，新一代显卡将始终拥有更卓越的“神经处理能力”，导致针对特定硬件量身定制的不同“变体”DLSS 5，这可能会让老旧显卡落伍。 对话还谈到了《侠盗猎车手6》的发布，尽管其预算巨大且采用手工打造的方式，但它与DLSS 5可能让其他资源较少的开发商“偷懒”的看法形成了对比。 展望未来，Linus认为，尽管英特尔和AMD最终会开发出竞争技术，但应用于网络摄像头的泄露版DLSS 5并不是“真正的”DLSS 5，因为它缺乏准确物体预测所需的关键矢量数据，导致“奇怪的故障”。然而，Linus和Gautier都总结道，这已经是“你将看到的最差版本的DLSS 5”。他们预计它将显著改进，变得“越来越常态化”，最终，AI生成内容与原生渲染内容将更难区分。

On the latest episode of The Vergecast, host David Amell welcomed Linus Sebastian and David Gautier from Linus Tech Tips to discuss Nvidia's recently released DLSS 5 technology, which dropped on September 3rd. The trio delved into why this new iteration of deep learning supersampling has sparked significant controversy within the gaming and developer communities. Historically, DLSS aimed to upscale lower-resolution renders using machine learning to fill in missing pixels, much like enhancing a low-resolution photograph. However, Linus Sebastian explained that DLSS 5 goes beyond this, "instead of just, you know, adding pixels that weren't there, it seems to be adding elements of the scene that weren't there or adding colors that weren't there." Nvidia has been explicit in stating their belief that "Moore's law... is completely dead," and that this "neural rendering" is the future. They envision a world where developers could simply command "barrel," and a photo-realistic barrel would be generated instantly and cheaply, without rendering all the pixels conventionally. However, David Gautier's hands-on experience with DLSS 5 on various games, including "Expedition 33" and a "Final Fantasy" title, paints a different picture. While acknowledging Nvidia's vision, he noted that the current implementation is far from the "magic barrel" ideal. Gautier likened it more to an "Instagram filter," emphasizing that its biggest limitation is the need to render 60 times a second, which restricts the AI's "creativity." He observed that good input yields better output, but it still retains an "AI look" that can elicit a "disgust reaction" from those averse to AI-generated content. The discussion quickly turned to the impact on artistic intent. Amell highlighted Gautier's observation that DLSS 5 could transform a stylized anime character into a "real person," potentially stripping away a game's unique aesthetic. The current controversy is complicated by the fact that the DLSS 5 DLLs being used by testers were leaked; only "NBA 2K27" has been officially tuned by Nvidia. Linus believes the technology will likely only benefit games striving for photorealism, while stylized games suffer. Gautier also noted a tendency for DLSS 5 to "suck the color out of it," akin to "lazy color correction in film." Some developers, like the creative director of "Kingdom Come Deliverance 2," have defended it, claiming it fulfills their original artistic vision that lacked the resources to achieve. Performance and cost were major points of contention. DLSS 5 incurs a "catastrophic" performance hit, even on a high-end 5090 GPU, with early demos reportedly requiring two 5090s. Linus questioned the value proposition, noting the "insane" cost of upgrading hardware (5080s at $1,500-$1,600). He suggested Nvidia's messaging is confusing, pushing high frame rates, resolutions, ray tracing, and AI enhancements simultaneously, which current hardware struggles to deliver. Gautier argued that Nvidia intentionally makes DLSS 5 demanding to ensure users require its performance uplift, thereby reinforcing dependence on their ecosystem. This forces PC gamers, accustomed to maxing settings on high-end hardware, into compromises. Both guests acknowledged Nvidia is reacting to the backlash, evidenced by plans for a "4000 series version of DLSS 5." While they expect improvements, Linus warned that newer generations of cards will always have superior "neural capabilities," leading to different "variants" of DLSS 5 tailored for specific hardware, potentially leaving older cards behind. The conversation also touched upon the "GTA 6" drop, which, despite its massive budget and handcrafted approach, serves as a counterpoint to the perceived "laziness" that DLSS 5 might enable in other studios with fewer resources. Looking ahead, Linus believes that while Intel and AMD will eventually develop competing technologies, the leaked DLSS 5 applied to webcams is not "true" DLSS 5, as it lacks the critical vector data needed for accurate object prediction, leading to "weird glitches." However, both Linus and Gautier concluded that this is "the worst version of DLSS 5 that you'll ever see." They anticipate it will significantly improve, become "more and more normalized," and eventually, it will be harder to distinguish between AI-generated and natively rendered content.