DLSS 5 ships: one-step pixel diffusion generates photorealistic 4K frames in real time

ctnzr · x · 2026-09-04

DLSS 5 is officially out: the first DLSS with a generative stage, using a real-time, 3D-guided one-step pixel-space diffusion model to turn rendered frames into photorealistic 4K scenes while preserving developer-authored content. The team explains why generation is needed — VRAM and compute cap both scene abstraction and ray counts, so even perfect reconstruction was bounded; DLSS 5 generates beyond that ceiling while staying grounded in the game and artist-controlled. Compared with offline models like GPT Image 2 and Gemini 3 Pro, it matches realism with far more faithful preservation of identity and scene structure. A technical report covers comparisons with offline generative models at similar scaling cost.

Original post →

More from Multimodal

Multimodal channel →