Intel is working on the holy grail of frame generation: AI-generated and no input latency

'GFFE: G-buffer Free Frame Extrapolation for Low-latency Real-time Rendering' is a new research paper published by Intel engineers on GitHub. The detailed report describes how Intel is looking to create a version of frame generation, as seen in DLSS 3, FSR 3, and Intel XeSS 2, "without introducing additional latency."

2

Sony's Horizon Forbidden West supports DLSS 3 Frame Generation, image credit: PlayStation Studios.

Technology like DLSS 3, a GeForce RTX 40 Series game changer, uses AI hardware to generate new frames and boost in-game performance. However, frame generation can cause noticeable lag without the latency reduction that naturally comes from rendering a game natively at 150 FPS.

This is why DLSS 3 works alongside NVIDIA Reflex to help improve latency and deliver a smoother, more responsive experience. Intel's new approach to frame generation is slightly different because it is not precisely frame generation but frame extrapolation.

At its most basic level, DLSS 3 Frame Generation takes data from two previously rendered frames, analyzes it, and then uses a sophisticated AI model to create (see: generate) a new frame and insert it between the two business-as-usual frames. Even though the result isn't perfect, the generated frame is only on screen for a short time, making it hard to spot any glaring difference.

The cost is latency; again, Reflex mitigates that when it comes to DLSS 3. Frame extrapolation is different because it uses a rendered frame history to generate new ones. Without looking at different frames to develop a new one and then inserting it into the rendering pipeline, frame extrapolation can sit directly in the pipeline without the need to access previous frames and things like motion vector data.

Of course, with less "data," quality becomes the stumbling block, which is where AI comes in. Extrapolation is not a new technology; however, Intel's GFFE method apparently "achieves comparable or better results than previous interpolation and G-buffer dependent extrapolation methods, with more efficient performance and easier integration." And it's not that far off from becoming a reality.