过去十几年,笔者一直在靠更精细的资产、更复杂的光照和更强的 GPU,一点一点把游戏推向真实。
今天的游戏之所以能做到这样的画面,本身就是模型、材质、动画、光照和算力不断堆叠的结果。

而 DLSS 5 带来的,是另一种格外直接的变化。
打开它之后,整个画面的光影关系会更自然,材质会更接近真实世界,细节之间也更容易形成统一的视觉关系。皮肤、织物、头发、金属、木地板这些原本很容易暴露「游戏感」的东西,会同时变得更可信。

很难想象,一项几乎发生在游戏渲染末端的图形工艺,能让游戏画面的真实性直接往前走一大步。
DLSS 5,让《NBA 2K27》接近次世代质感
作为一个从《NBA 2K9》开始,到现在几乎每一部都玩过几百个小时的玩家,我基本是亲眼目睹了 2K 这些年画面逐步变好的全过程。

先是《NBA 2K14》的 PS4 版本整套人物和球场资产几乎重新制作,球员的模型、面部、灯光和地板反射,相比 PS3 时代有格外显著的跨越。
再往后到《NBA 2K17》,PS4 这一代的画面基本进入成熟期。人物扫描越来越准确,灯光、球馆和转播包装越来越完整,这也是后来很多代作品一直沿用的视觉基础。
来到 PS5 世代,《NBA 2K21》则又一次定义了篮球游戏的标杆,材质、灯光、皮肤 Shader、汗水、球衣和球场细节又总体往前走了一步。

今天《NBA 2K27》的基础画面,某种程度上仍旧建立在这一轮升级之上。
把十几年的作品放在一起看,进步当然格外显著。
难题是,这种进步也来得越来越慢。

即使到了《NBA 2K26》,甚至今年的《NBA 2K27》,画面已经足够精细,但你还是很容易一眼看出它是一款游戏。
最显著的就是那种 3D 模型感。
球员的脸扫描得很准,毛孔、汗水、胡须都有,但皮肤仍旧容易显得像塑料;明暗关系比较简单,衣服也经常有一种「贴在模型表面」的感觉。整个画面的颜色、材质和光线,仍旧带着一种体育游戏特有的卡通感。
但一项神奇工艺的浮现,改变了这种现状。

当我今年第一次打开 DLSS 5 ,看到卢卡·东契奇的面部,就感觉一切都不一样了。光线在脸上的过渡显著更复杂。眼窝、鼻翼、嘴唇、耳朵这些地方都开始浮现现实人脸里格外细微的明暗和透光关系。
一瞬间,横档在面前的文班亚马,也突然没有那么像一个 3D 模型了。
这种真实性也会浮现在大量以前容易被忽略的细节里。
例如球场地板。

木纹、划痕、磨损、反光这些东西本来一直都有,但 DLSS 5 会让这些细节之间的关系更自然。你看到的不再只是一张高清的木地板材质,而更像一块真正经历过比赛、灯光照射和反复摩擦的地板。

球衣一直是过往游戏里格外容易暴露「假」的东西,因为现实中的布料有极其复杂的纤维、褶皱、透光和反射。
打开 DLSS 5 之后,球衣表面的织物感显著增强,褶皱里的明暗更细,衣服不同角度接受光线时的变化也更接近真正的布料。
就连整个画面的色彩和光影,也有了很好的修正。
过去 2K 的画面一直比较鲜艳,人物、球衣、地板和球馆之间有一种很显著的「游戏色彩」。尤其是洛杉矶湖人队的球场,总有一种过曝的感觉。

DLSS 5 会把这种感觉显著压下去。
颜色更加克制,光线更加自然,不同材质之间开始具备更接近现实世界的关系。
综合来看,当球员、衣服、地板、灯光和色彩一起变化之后,整个画面开始从「高质量的 3D 游戏」,无限接近了真实摄影机拍出来的 NBA 比赛质感。
画面进化的第三条路:把真实感「推断出来」
思考这样一个难题:如何让游戏画面无限接近真实世界?
做更高精度的模型、更复杂的游戏资产?再让 GPU 去计算更多光源、反射、全局光照和各种图形成效?
简单来说,就是开发者把游戏世界做得越来越复杂,再由 GPU 把这个世界算出来。
这条路线当然有效,但发展到今天,边际收益已经越来越低。
《最后生还者 2》是 2020 年、PS4 世代末期的游戏,但今天重新去看,它的人物、材质、动画和总体光影表现,仍旧比很多 PS5 世代的作品更加出色。

继续向真实逼近,需要付出的资产成本和计算量越来越高,最终可以被玩家直接感受到的变化却越来越小。
所以到了 DLSS 5,NVIDIA 挑选了另一条路:
直接从现实世界里学习,现实究竟应该长什么样。
从第一性原理去理解 DLSS 5,实际上格外简单:
与其花费大量时间去制作资产,大量算力计算优质画面,那还不如在已经渲染的画面上,二次加工。
因为 AI 已经看过了现实世界,它可以直接参与最终画面的生成。用这种取巧的方式,取得视觉上的真实性。

当游戏引擎和 GPU 已经完成基础渲染,得到一张游戏画面之后,DLSS 5 会在整个图形管线格外靠后的阶段,再用一个神经网络对这张画面开展生成式渲染。
NVIDIA 把这套工艺叫做 3D-Guided Neural Rendering。
它拿到的不只有当前已经渲染好的 RGB 画面,同时还有 Motion Vectors、上一帧留下的 Temporal State,以及开发者给出的 Masks 和艺术控制信息。
这些来自 3D 游戏世界的数据负责告诉 AI:
画面里究竟是什么。
而神经网络从大量真实视觉数据中学到的 Appearance Prior,则负责告诉它:
这些东西在现实世界里通常应该长什么样。
整个过程要塞进一帧十几毫秒的实时预算里,神经渲染通常只能占其中几毫秒。这也是它使用单步模型的原因。然后,只用几毫秒完成这次推理,生成一张符合现实世界的真实感画面,最终展现在玩家眼前。

所以它并不是简单给游戏套一层滤镜。
AI 已经看过现实世界,它可以直接参与最终画面的生成。
这也是为什么我会觉得,DLSS 5 是一项真正具有跨世代意义的工艺。
一旦理解它到底在做什么,就会发现,它改变的并不只是《NBA 2K27》的画面,而是整个游戏行业生产画面的方式。
跨越世代,真实画面不再依靠手搓
当然,今天的 DLSS 5 还远远谈不上成熟。
最直接的难题就是表现。
复杂的神经网络运算,意味着它仍有显著表现成本,当下在《NBA 2K27》中也有 RTX 50 系列的硬件门槛。
在我这台 RTX 5080 上,《NBA 2K27》4K 最高画质关闭 DLSS 5 时,游戏可以跑到 120 帧;打开之后,则会直接掉到 45 帧左右。可见,当下这项工艺还有相当大的优化空间。

它也并不合适所有游戏。
《NBA 2K27》这样的体育游戏本就追求写实,会最先从 DLSS 5 中受益。
但《塞尔达》《马里奥》这类作品追求的是自己独特的视觉语言。对于它们来说,让 AI 学习「现实世界长什么样」,意义就要小得多。将来如果神经渲染真正走向普及,它需要学习的可能不只有现实,还涵盖不同游戏自己的艺术风格。
而另一个更关键的难题,是真实和忠实之间的边界。
早期 DLSS 5 的演示之所以引发争议,就是因为 AI 可以让一张脸变得更加真实,同时也有可能让它偏离开发者原本外观的那张脸。

不过这种争议,在玩家们看到 DLSS 5 对各种游戏带来的画面革新之后,仿佛一下子烟消云散——
那些熟悉的游戏,有些甚至已经发售接近二十年之久,在 DLSS 5 的加持之下,秒变画面重制版,整个画面表现力焕然一新,甚至让人有重新玩一遍的冲动。
这足以说明,DLSS 5 是一项值得期待的工艺。

回看过去游戏开发的路径,模型越复杂、资产越精细、光照越真实,制作成本就越高,对玩家电脑硬件的压力也水涨船高。这套方法已经越来越接近一堵无法逾越的成本之墙。
DLSS 5 打破了最终视觉复杂度必须和手工制作复杂度同步增长的铁律。当一部分原本需要被建模、绘制和计算的视觉细节,可以直接由神经网络从现实世界的经验中生成,它改变的实际上是整个游戏行业生产画面的方式。
而作为第一款真正用上 DLSS 5 的游戏,《NBA 2K27》也因而站在了一个全新时代的起点之上。
本文来源:https://www.ifanr.com/1682974
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Over the past decade or more, I have been gradually pushing games toward realism through more detailed assets, more complex lighting, and more powerful GPUs.
The reason today's games can achieve this kind of image quality is the result of continuously stacking models, materials, animation, lighting, and computing power.

What DLSS 5 brings is a different kind of especially direct change.
Once it is turned on, the lighting and shadow relationships across the entire image become more natural, materials become closer to the real world, and details more easily form a unified visual relationship. Skin, fabric, hair, metal, and wooden floors—elements that once easily gave away the 'game-like' feel—simultaneously become more believable.

It is hard to imagine that a graphics technique occurring almost at the end of a game's rendering pipeline could push the realism of game visuals forward by a huge step.
DLSS 5 brings NBA 2K27 close to a next-generation feel
As a player who has started with NBA 2K9 and played hundreds of hours of almost every installment since, I have essentially witnessed the entire process of 2K's visuals gradually improving over the years.

First, the PS4 version of NBA 2K14 nearly rebuilt its entire character and court assets. Player models, faces, lighting, and floor reflections showed an especially significant leap compared with the PS3 era.
Later, by NBA 2K17, the visuals of the PS4 generation had basically entered a mature stage. Character scans became increasingly accurate, while lighting, arenas, and broadcast presentation became increasingly complete. This also became the visual foundation that many later installments continued to use.
With the PS5 generation, NBA 2K21 once again set the benchmark for basketball games. Materials, lighting, skin shaders, sweat, jerseys, and court details all moved forward overall.

Today, the base visuals of NBA 2K27 are still, in a certain sense, built on this round of upgrades.
Looking at the series across its dozen or so years, the progress is remarkably striking.
The problem is that this progress is also becoming slower and slower.

Even in NBA 2K26 and even this year's NBA 2K27, the graphics are detailed enough, yet it's still very easy to tell at a glance that it's a game.
The most obvious giveaway is that 3D-model feel.
Players' faces are scanned with high accuracy—pores, sweat, stubble, everything—but the skin still tends to look plastic. The lighting and shading relationships are too simple, and clothing often feels like it's merely "stuck onto the surface of the model." The overall colors, materials, and lighting still carry that distinctive cartoonish quality found in sports games.
But the emergence of one remarkable technology has changed this landscape.

Image|Character face, DLSS 5 ON/OFF
When I first booted up DLSS 5 this year and saw Luka Dončić's face, I felt everything was suddenly different. The light transitions on the face became significantly more complex. The subtle interplay of shadows and translucency around the eye sockets, nostrils, lips, and ears began to resemble the finest details of a real human face.
In an instant, the Wembanyama standing across from me no longer looked quite so much like a 3D model.
This sense of realism also emerges in countless details that were previously easy to overlook.
For example, the court floor.

Image|Floor details
The wood grain, scratches, wear, and reflections were already present, but DLSS 5 makes the relationships between these details far more natural. What you see is no longer just a high-definition wooden floor texture, but something that looks more like a surface that has genuinely endured games, arena lighting, and repeated friction over time.

Image | Jerseys with DLSS 5 On / Off
Jerseys have always been one of the most obvious telltale signs of artificiality in games, because real-world fabrics have incredibly complex fibers, wrinkles, translucency, and reflections.
With DLSS 5 enabled, the woven texture on jersey surfaces becomes noticeably more pronounced, the shading inside folds grows finer, and the way garments respond to light from different angles more closely resembles actual fabric.
Even the overall color and lighting of the scene receives solid correction.
In past 2K titles, the visuals tended to be overly saturated, with a very noticeable "gamey" color palette across players, jerseys, the court floor, and the arena. The Los Angeles Lakers' home court, in particular, always had an overexposed look.

DLSS 5 will significantly reduce this feeling.
Colors become more restrained, lighting more natural, and the relationships between different materials begin to resemble those in the real world.
Taken together, as players, uniforms, floors, lighting, and colors all change, the entire scene begins to shift from "high-quality 3D game" visuals to something almost indistinguishable from real-camera NBA broadcast footage.
The third path of visual evolution: inferring realism
Consider this challenge: how do you make game visuals approach the real world infinitely closely?
Build higher-precision models and more complex game assets? Have the GPU calculate more light sources, reflections, global illumination, and various rendering effects?
In simple terms, it means developers make the game world progressively more complex, and the GPU computes that world.
This approach is certainly effective, but as we've reached today, the marginal returns have been diminishing.
The Last of Us Part II, released in 2020 at the tail end of the PS4 generation, still looks superior in its characters, materials, animations, and overall lighting compared to many PS5-generation titles when revisited today.

Continuing to push closer to realism requires ever-increasing asset costs and computational power, yet the changes directly noticeable to players keep getting smaller. So with DLSS 5, NVIDIA chose a different path: learning directly from the real world what reality should actually look like. From first principles, understanding DLSS 5 is surprisingly simple: instead of spending massive amounts of time creating assets and heavy compute to produce high-quality images, it's more efficient to perform secondary processing on already-rendered frames. Because AI has already seen the real world, it can directly participate in generating the final image. Through this clever approach, it achieves visual realism.

Once the game engine and GPU have completed their base rendering to produce a frame, DLSS 5 applies generative rendering through a neural network at a very late stage in the graphics pipeline. NVIDIA calls this process 3D-Guided Neural Rendering. It receives not only the currently rendered RGB frame, but also Motion Vectors, the Temporal State from the previous frame, as well as Masks and artistic control information provided by developers. These data points from the 3D game world tell the AI what is actually in the scene. Meanwhile, the Appearance Prior that the neural network has learned from vast amounts of real visual data tells it what those things typically look like in the real world. The entire process must fit within a real-time budget of roughly ten-plus milliseconds per frame, with neural rendering consuming only a few of those milliseconds. This is why it uses a single-step model. In just a few milliseconds, inference is completed to generate a photorealistic frame that reflects the real world, ultimately presented to the player.

So it is not simply slapping a filter layer over the game. AI has already seen the real world, and it can directly participate in generating the final image. This is why I consider DLSS 5 a truly cross-generational technology. Once you understand what it actually does, you realize it changes not just the visuals of NBA 2K27, but the entire way the gaming industry produces images across generations—realistic visuals no longer depend solely on manual craftsmanship. That said, today's DLSS 5 is far from mature. The most immediate challenge is performance. Complex neural network computations mean it still carries significant performance costs, and currently requires RTX 50 series hardware for NBA 2K27. On my RTX 5080, NBA 2K27 at 4K highest settings runs at 120 fps with DLSS 5 turned off, but drops to around 45 fps with it enabled. This clearly shows there is still considerable room for optimization.

It is also not suitable for all games. Sports titles like NBA 2K27, which inherently strive for realism, will benefit the most from DLSS 5. But games like The Legend of Zelda and Mario pursue their own unique visual language. For them, having AI learn what the real world looks like is far less meaningful. In the future, if neural rendering truly becomes widespread, it may need to learn not only reality but also the distinct artistic styles of different games. Another more critical challenge lies in the boundary between realism and faithfulness. The early DLSS 5 demos sparked controversy precisely because AI could make a face look more realistic while potentially straying from the developer's intended appearance.

However, this controversy seemed to vanish almost overnight once players saw the visual transformations DLSS 5 brings to various games—those familiar titles, some of which have been released nearly two decades ago, instantly become visually remastered under DLSS 5's boost. The overall visual presentation is completely revitalized, even inspiring players to replay them. This is sufficient proof that DLSS 5 is a technology worth looking forward to.

Looking back at the trajectory of game development, as models become more complex, assets more detailed, and lighting more realistic, production costs rise accordingly, and the pressure on players' PC hardware climbs with them. This approach is increasingly approaching an insurmountable wall of cost. DLSS 5 breaks the iron law that final visual complexity must grow in lockstep with manual production complexity. When a portion of the visual details that once needed to be modeled, painted, and calculated can now be generated directly by neural networks from their experience of the real world, what it fundamentally changes is the entire way the gaming industry produces images. And as the first true DLSS 5 title, NBA 2K27 stands at the threshold of a brand-new era.