DLSS 5 launches on September 3: NVIDIA turns the upscaler into the final stage of rendering

NVIDIA has closed the main question around DLSS 5: the technology becomes available on September 3 at 9:00 PM Pacific time, NBA 2K27 has been chosen as the debut game, and support covers all GeForce RTX 50 Series graphics cards and laptops, as well as GeForce NOW. The announcement says nothing about RTX 40.

The date is not the main thing here. DLSS is ceasing to be merely a reconstruction technology: earlier versions completed what the engine had already computed — resolution, additional frames, clean ray-traced lighting. DLSS 5 with 3D-Guided Neural Rendering adds what was not in the frame at all: details of lighting and materials that developers gave up for the sake of real time. For artists and technical directors this changes the structure of the pipeline — the final look of the image is formed after the engine has handed over the frame.

The engine frame remains the law for the neural network

NVIDIA draws a line between its technology and generation from a text description. The basis is a rendered frame with all the authored geometry, textures and lighting buffers, while the neural model works on top of it as the last stage of the graphics pipeline. The company’s wording: the engine frame determines what has to remain unchanged, and the artist sets what the neural network is allowed to change. DLSS 5 remains optional and can be switched off at any moment.

The complaint that has drawn criticism to the technology since its spring premiere concerned precisely this boundary: after processing, characters’ faces often looked like images generated by a neural network. NVIDIA head Jensen Huang defended the development, stating that he himself does not like “AI slop”. The set of manual controls in the release version looks like an answer to that criticism.

The artist decides what the neural network must not touch

Developers have been given five layers of control over the effect:

• Model selection — several sets of weights with different training parameters, each giving the final image its own character;
• Structure Intensity — the strength of high-frequency detail: ambient occlusion, contact shadows, subsurface scattering;
• Tone Intensity — the low-frequency component: lighting and colour response;
• Automatic masking — the model itself picks out objects in the frame, characters for example, and is applied to them selectively, without manual markup;
• Developer masking — per-pixel masks for specific objects.

The practical point of this set is tied to art direction: masks make it possible to confine the neural layer to specific objects, while automatic selection, according to NVIDIA’s description, relieves the team of tagging objects by hand.

The picture has to be paid for in frames

Performance has been the weak spot of DLSS 5 from the very beginning. NVIDIA ran the first demonstration on two GeForce RTX 5090s; now a single card is enough, and the company speaks of a fivefold performance gain over six months. But its own measurements were made with Multi Frame Generation enabled in 6X mode — hence the 233 and 370 frames per second the company quotes.

If the sixfold frame generation is taken out of those numbers, the native rendering rate turns out to be very low — in a number of scenarios noticeably below 60 frames per second. NVIDIA has not published a direct “DLSS 5 on / off” comparison, so the real cost of the effect cannot be assessed before independent tests. According to the available information, optimisation work will continue until the end of the year.

A debut in a basketball simulator

The choice of NBA 2K27 for the premiere looks less than obvious: the series has never been regarded as a showcase for graphics technologies. At the same time, a sports simulator gives a technology that deals with skin, fabric and light plenty of close-ups of players’ faces. Besides local RTX 50 cards, the technology is available in the GeForce NOW cloud. The details are explained by the developers and the authors of the technology themselves.

Why NVIDIA is moving rendering into a neural network right now

The absence of RTX 40 from the support list — against the background of modders having got DLSS 5 running on older cards over the weekend with roughly the same performance drop — reads more like a product decision than a hard technical limitation. That guarantees nothing to owners of the previous generation, but after such experiments the argument that “the hardware cannot handle it” sounds weaker.

The very appearance of 3D-Guided Neural Rendering fits a trend that stretches back to the first versions of DLSS: the share of the frame computed in the traditional way is consistently shrinking, while the share restored or completed by the network is growing. First the neural network was responsible for resolution, then for additional frames and the reconstruction of ray-traced lighting, and now for lighting and materials. If the trend continues, the raster and ray-traced passes will gradually turn into a detailed set of hints for the neural model, and the chain “artist — shader — engine” will gain one more link, whose behaviour is tuned with sliders and masks rather than with code.

For studios this involves quite specific risks. The final look of a game starts to depend on the model version on the driver side, which means art direction will have to be checked not only in the editor but also at the output of the neural layer. For users the question is simpler: as long as the declared effect rests on frame generation in 6X mode, a conversation about image quality runs into a conversation about input latency. The first independent measurements of DLSS 5 will show how expensive the declared authorial vision is in practice.

Why the bet on painting in the frame is being made right now: the limit of rasterisation, the promotion of RTX 50, the race for FPS numbers?