Nvidia brought DLSS 5 back to the stage at SIGGRAPH 2026 in Los Angeles on July 21, hoping a reworked demo would calm the anger that followed its first showing. The technology looks cleaner than before, and it runs in real time, but the core complaint hasn’t gone away. A lot of people still think it turns game characters into AI-smoothed versions of themselves.
Quick answer: DLSS 5 is Nvidia’s neural rendering system that adds photorealistic lighting and material detail on top of a game’s existing graphics. The SIGGRAPH build adds three developer-selectable AI models, per-scene and per-character controls, and a claimed sub-16ms render time at 4K. It is expected to launch this fall, with system requirements still unannounced.
What DLSS 5 actually does
DLSS 5 is the newest version of Nvidia’s Deep Learning Super Sampling suite. Older versions let a game run at a lower internal resolution and then rebuild the image to look sharper or run faster. DLSS 5 goes further. It uses a real-time neural rendering model that reads a frame’s color and motion data, then generates extra lighting, textures, hair, skin, and fabric detail on top of what the engine already drew.
Nvidia framed this as a bigger shift than a normal upscaler. When the company first unveiled DLSS 5 in March 2026, CEO Jensen Huang called it “the GPT moment for graphics” and described it as blending handcrafted rendering with generative AI. The idea is to keep the artist’s structured 3D scene as the ground truth, then layer realistic materials over it.
Three DLSS 5 models: A, B, and C
The most concrete change from the earlier demo is choice. Developers can now pick from three DLSS 5 models, labeled Model A, Model B, and Model C. Each one applies a different level of structural intensity, global illumination, and texture detail, and each carries a different parameter count. A heavier model uses more compute and memory than a lighter one, but Nvidia says all three run on a single GPU.
The models aren’t locked to a whole game. A studio can mix and match them across different scenes or characters, and the upscaler can switch between models in real time depending on what a scene needs.
| Aspect | Detail |
|---|---|
| Models offered | Model A, Model B, Model C |
| What they change | Structural intensity, global illumination, texture detail, and other visual effects |
| Parameter count | Different per model; heavier models use more compute and memory |
| Hardware | All variants run on a single GPU |
| Control level | Adjustable per scene and per character |
Keeping the artist’s intent
The loudest complaint about the first demo was that DLSS 5 overrode the original art. Nvidia responded by adding fine control. Developers can now configure how individual objects, scenes, and settings are processed, keeping control over character appearance, lighting, animation, and overall art direction.
Nvidia says the hard part is holding onto visual identity, object semantics, character poses, output lighting, motion, and art direction while still adding photorealistic materials. Inside those limits, the system applies its material realism without changing the underlying geometry. Digital Foundry noted that none of the original assets are touched. The main thing DLSS 5 alters is how light hits the scene.
How it runs in real time
Normal generative AI video tools build several frames at once, reasoning across a chunk of footage. Games can’t wait for that. DLSS 5 generates images one frame at a time using motion vectors pulled straight from the game engine. Because the model knows exactly where objects are moving, it can attach new detail instantly, which Nvidia says avoids shimmer, flicker, and image corruption.
Speed is the other headline. Nvidia states that rendering a single 4K frame, which is 8.3 million pixels, takes less than 16 milliseconds, enough for more than 60 frames per second. Most of that time is spent by the game engine itself rather than the DLSS 5 model. The company also stresses that the model is memory efficient, though it hasn’t said how much VRAM comfortable 4K gaming will actually need.
| Metric | Nvidia’s claim |
|---|---|
| 4K frame time | Under 16 ms per frame |
| Resolution | 4K (8.3 million pixels) |
| Frame rate | Above 60 FPS |
| Generation method | Frame by frame, using engine motion vectors |
| GPU count | Single GPU |
Why people still aren’t convinced
The reworked demo lands better than the March version, which drew almost entirely negative reactions and a wave of memes after screenshots showed characters like Resident Evil Requiem‘s Grace looking noticeably smoothed and made up. This time the effect looks less like a phone filter, and some observers admitted it is a real step up. But the fundamental objection remains.
In the SIGGRAPH footage, dragging sliders for “Structure Integrity” and “Tone Intensity” visibly changed a character’s facial texture. To critics, that is exactly the problem. If every face ends up with the same glossy AI skin, then the result reads as a beauty filter rather than better graphics.
Others question the premise itself. The argument is that upscaling should exist to make a game run faster, not to add or rewrite what the artists already made.
How much of this shows up in shipped games is up to developers, since they control which model runs and how aggressively it is applied. A studio can leave a scene untouched or push the effect hard. That flexibility hasn’t stopped the “AI slop filter” label from sticking online.
When DLSS 5 launches
DLSS 5 is expected to launch this fall. Nvidia says it will share general system requirements and more technical detail closer to release, including how the VRAM story shakes out. Until then, the open question is whether 4K neural rendering will be limited to high-end cards or reach a wider range of GPUs.
That timing sits behind a smaller update. DLSS 4.5, which focuses on ray reconstruction, is set to arrive in August and is a more conventional step than the neural rendering push. For now, DLSS 5 remains a demo that impresses on the technical side while leaving many players unsure they want it switched on.