NVIDIA's Neural Texture Compression emerges as a key new component in AI graphics

  • Neural Texture Compression (NTC) reduces VRAM usage by up to seven times while maintaining visual quality.
  • The demo "Tuscan Wheels / Tuscan Villa" shows scenes that go from 6,5 GB to about 970 MB of memory.
  • "Neural Materials" allow fewer material channels and speed up rendering by between 1,4 and 7,7 times.
  • NTC is integrated into the graphics pipeline along with technologies such as DLSS to optimize memory and fidelity.

NVIDIA Neural Texture Compression Technology

In recent years, there has been much talk about 4K resolutions, ray tracing, and DLSS, but one of the major, often overlooked limitations remained graphics memory. With increasingly heavy textures and more demanding games , even many mid-range graphics cards in Europe quickly ran out of VRAM, forcing players to lower quality settings or endure stuttering and performance drops.

In this context, NVIDIA's Neural Texture Compression (NTC) emerges , a texture compression technology based on neural networks presented in detail during the 2026 GTC and GDC. Its proposal is straightforward: to very aggressively reduce the memory consumption of textures, but without worsening the image perceived on screen, and even improving it in some scenarios compared to classic methods.

How Neural Texture Compression works and what makes it different

NTC's foundation lies in the use of small neural networks specifically trained for textures , rather than relying solely on the block compression formats (BCn, BC6, BC7, etc.) that the industry has been using for years. Traditionally, these textures are stored already compressed in VRAM and interpreted on the fly by the GPU, but they still occupy a significant amount of memory.

With Neural Texture Compression, texture information is stored in a much more compact representation , a kind of latent encoding that the neural network decodes in real time when rendering each frame. Instead of handling gigabytes of diffusion maps, normals, roughness, and so on, the GPU works with a significantly smaller data package.

According to NVIDIA, these neural models have been trained to understand what a texel (the "pixel" of a texture) should look like for a given material: stone, wood, metal texture , ceramic, fabrics, etc. Based on this learning, the network can reconstruct the final appearance from the compressed data, emulating the visual result we would have with much larger textures.

The practical result is that textures cease to be a fixed "burden" in memory and instead depend on smaller compressed code and the inference capabilities of AI , something that fits with NVIDIA's general trend of shifting memory and bandwidth load towards intelligent computing in its GPUs.

The demo "Tuscan Wheels / Tuscan Villa": from 6,5 GB to less than 1 GB of VRAM

To illustrate the potential of the technology, NVIDIA has shown several technical demonstrations, including the already well-known "Tuscan Wheels" or "Tuscan Villa" scene , a Mediterranean villa environment with detailed interiors that serves as a high-level-of-detail testing ground.

In the traditional configuration, using standard BCn formats, the scene requires around 6,5 GB of VRAM for textures alone. By enabling Neural Texture Compression, the same environment runs with approximately 970 MB of graphics memory , representing a reduction of nearly 85% compared to the original usage. Similar figures have been seen in other demonstrations, with reductions of around 80% (down to about 670 MB in certain scene variations).

The key difference isn't just the magnitude of the VRAM reduction, but that the visual comparison between the two versions is virtually indistinguishable to the average user. According to NVIDIA, if the same memory budget is maintained, NTC is even capable of preserving more fine detail than downscaled or resized BCN textures.

A concrete example can be seen inside the villa, with a table covered in dishes, bottles, and decorative objects. In this comparison, the section processed with NTC and the same amount of memory shows greater sharpness and micro-detail than the section rendered with reduced BCN textures to fit the same VRAM budget.

These types of results suggest the dual use of technology: studios can choose to drastically reduce memory consumption without losing quality , or maintain consumption and raise the level of visual fidelity, something especially attractive for projects that aspire to an almost photorealistic look.

Practical advantages for games and graphics engines

From a development perspective, the main consequence of all this is that textures are no longer so restrictive when designing complex scenes . Less VRAM occupied by surface maps means more headroom for other systems or for increasing the amount of content displayed simultaneously.

For PC games aimed at the European market, where a significant portion of the user base still uses mid-range GPUs or GPUs with 8 GB of memory , such aggressive compression opens the door to enabling high-resolution textures on systems that, currently, are forced to cut quality to avoid exceeding the VRAM limit.

Among the advantages that NVIDIA has highlighted are:

  • Drastic reduction in VRAM usage, up to seven times in specific scenarios.
  • Ability to drive higher resolution textures without triggering memory requirements.
  • Bottleneck relief related to memory bandwidth and asset streaming.
  • Potential smaller installation and patch size, by packaging more compact textures onto disk.
  • Better use of portable devices and future consoleswhere memory is a more limited resource.

This aligns with a market where, even in Europe, games easily exceed 100 GB in download size and where available bandwidth isn't always sufficient, especially in rural areas or with less powerful connections. Reducing texture size without sacrificing quality can make a significant difference in download and update times.

Furthermore, by relying more on intelligent computing, some of the pressure on physical memory is relieved and the computing capabilities of modern GPUs are used more efficiently, something that NVIDIA has been promoting for some time with other AI solutions.

Neural Materials: fewer channels, more speed

Alongside NTC, NVIDIA has also introduced the concept of Neural Materials , a natural extension of the idea of ​​neural compression applied not only to textures, but to the physical model of the materials used in rendering.

In a traditional workflow, describing how a surface behaves in response to light involves combining multiple channels and maps : base color, normals, roughness, metallicity, occlusion, and other specific data linked to the BRDF equation used by the graphics engine. This results in a large amount of data, numerous memory accesses, and a significant number of mathematical operations per pixel.

With Neural Materials, that set of channels is reduced to a more compact latent representation that a small neural network is responsible for decoding in real time, reconstructing the visual properties of the material during rendering.

In tests shared by NVIDIA, a 19-channel material configuration was reduced to just 8 channels , which, in 1080p resolution test scenes, resulted in accelerations of between 1,4 and 7,7 times in rendering time, depending on the specific case.

This approach not only saves memory, but also simplifies the number of data accesses and operations per pixel , something especially critical in ray tracing and path tracing configurations, where the cost of each light bounce is multiplied.

NTC within NVIDIA's new AI-powered graphics pipeline

Neural Texture Compression isn't coming alone. It's part of a broader strategy in which NVIDIA is redesigning key parts of the graphics pipeline using neural networks . Technologies like DLSS, frame generation, and the upcoming DLSS 5 are all based on the same underlying idea: shifting some of the traditionally "hard" rendering work to optimized AI models.

Several technical talks at the GTC have explained that, beyond simple image post-processing, compression and neural rendering of textures and materials are key components for this ecosystem to function properly. Reducing memory, minimizing bottlenecks, and freeing up resources allows other stages, such as DLSS image reconstruction, to have more headroom.

One point NVIDIA emphasizes is that, in the case of NTC and Neural Materials, generative AI isn't being used to "invent" artistic content , but rather inference models designed to faithfully reproduce the appearance of textures and materials previously defined by the art team. This aims to address some of the criticism that has arisen in the community regarding the potential impact of AI on the original visual intent of games.

In practice, the stated goal is for neural tools to act as technical accelerators , and not as substitutes for the creative work of artists and studios, something that will surely continue to generate debate as these solutions reach commercial titles.

According to NVIDIA, the networks that support NTC have already been trained for a wide variety of materials commonly found in video games , which in theory would facilitate their integration into commercial engines once the technology is opened to developers.

Potential impact on the European market and future deployment

To date, NVIDIA has not set a specific date for Neural Texture Compression to be widely implemented in commercial games, but technical demonstrations point to a scenario in which the use of VRAM could become less restrictive , especially on PC.

In Europe, where the hardware landscape is highly diverse, this type of solution could have a significant impact. Many users play on gaming laptops or desktops with GPUs containing 6 or 8 GB of VRAM , an amount that is already starting to fall short in some AAA releases with "ultra" textures. If NTC lives up to its promises, it could allow users to maintain these high settings without the game experiencing swapping or abrupt performance drops.

From the developers' perspective, there are also practical incentives: less texture data means lighter builds , smaller test downloads, and more reasonable update times. For European gamers, where not everyone has access to high-speed fiber optic connections, this could translate into a somewhat less frustrating experience when installing or updating large games.

However, there will also be factors to consider. The actual adoption of Neural Texture Compression will depend on the ease of integration into engines like Unreal Engine, Unity, or other internal engines , the support offered by different generations of GPUs, and the balance between quality, performance, and implementation cost for each studio.

In any case, what does seem clear is that graphics memory has become a priority target for optimization , and that proposals like NTC fit into a broader trend towards "neural rendering", where intelligent computation replaces some of the traditional solutions based solely on brute force.

Looking at the big picture, Neural Texture Compression, Neural Materials, and the other techniques announced around DLSS point to a generation of graphics engines where AI not only creates pixels but also decides how to store, compress, and reconstruct them . If the promises of VRAM savings, improved detail, and reduced rendering times materialize in commercial games, we could be witnessing one of the most significant changes in recent years in how real-time graphics are produced and executed.

Shading settings for matte materials: PBR guide for roughness, lighting, and textures
Related article:
Shading settings for matte materials: PBR guide for roughness, lighting, and textures

Add as preferred source in Google