Some of you are probably already almost knowledgeable about Neural Renderingbut, if this is not your case (yet), here is a video that might interest you. NVIDIA calls it “Introduction to Neural Rendering”, quite a program, therefore…! It was published on the firm’s YouTube channel today, April 4, 2026, very early in the night. The presentation that you will be able to consult was, however, actually made during theGTC 2026 from NVIDIA, in the middle of March.
The two speakers, Shannon Woods and Alexey Bekin, present to us for almost 40 minutes the challenges of switching to machine learning for real-time graphics rendering.
They first present to us the three “patterns”, which can also be mixed. ML can only intervene in post-processing, as for DLSS 5 for example. It can otherwise be integrated into the pipeline, to intervene directly in rendering. This is for example the case of Neural Materials or even Neural Texture Compression (NTC), the two main subjects that will be discussed in this presentation. The third pattern is the generative approach, with very reduced intervention (or even absence) of the pipeline, the neural network generating a large part (or all) of the image.
We then get to the heart of the matter, watch out for a headache if that’s not your cup of tea, with several minutes around the infrastructure necessary for real-time neural rendering. NVIDIA has chosen the shading language Slang a you binding Python SnakePywhich will undoubtedly be a good leg up for those who have two feet outside the world of development. You will at least be able to shine in society by giving a little change after seeing this…!
Then it’s time for case studies, which will undoubtedly be the most ambiguous and clear passages for many people. Alexey Bekin shows us comparisons of scenes in Block Compression n (BCn), the standard compression used for around twenty years, then in Neural Texture Compression. The first scene is intended to have “almost equivalent visual rendering” and would be around 6.5 GB in BNc compared to 970 MB in NTC. The second scene for its part aims for a size of 970 MB, both in BNc and in NTC, and NVIDIA insists on the superior quality of the textures in NTC. The conclusion is obvious… NTC allows you to have, for the same memory size, a better rendering, or to consider reducing the quantity of VRAM used while maintaining good visual quality.
top BNc, bottom NTC, Ã 970 Mo
Alexey Bekin then addresses Neural Materials and their interest. Managing the many materials that exist in the world is very complex to render properly for traditional shaders. Neural Materials can greatly improve rendering times, and NVIDIA presents as an example the case of a ceramic in Path Tracing (1080p) with an 8-channel neural version which allows rendering 1.4 to 7.7 times faster than traditional shaders (19 channels).
The last case study concerns less the PC world on the other hand, and focuses on Omniverse NuRec for training autonomous vehicles.
