Nvidia released studio driver 595.79, bringing targeted performance improvements across several AI inference formats and creative applications.

Key improvements

Release notes describe optimizations for NVFP4, NeuroStream, LTXV, Flux 2 Klein and WindowsML GPU inference in common toolchains and workflows.

• NVFP4 model optimizations reported for Comfy, aiming to improve throughput in compatible generative model pipelines.
• NeuroStream receives a performance boost in Topaz, with reported latency and throughput improvements for certain workloads.
• LTXV inference shows a 50% speedup when using FP8 weights and a 30% reduction in VRAM consumption in some scenarios.
• Flux 2 Klein can run up to twice as fast when using multiple references, and gains a 40% boost with FP8 weights plus 30% lower VRAM use.
• VoiceMod inference on WindowsML is offloaded to the GPU instead of the CPU, enabling hardware-accelerated audio processing on supported systems.

Compatibility and caveats

Manufacturers note that NVFP4 weight optimizations are limited to the 50th series of GPUs, so older or different models may not show the same benefits.

On a personal test note, a user with a GeForce 4090 indicated plans to prioritize FP8 and higher optimizations rather than NVFP4-specific improvements on their hardware.

User experience and testing

Previously, after installing a studio driver with a clean setup, the same user reported either no measurable improvement or reduced performance in Z-image workloads.

Following that experience they reverted to the default game driver, which receives more frequent updates and has been their stable baseline for testing and comparison.

The current release invites further validation across models and hardware configurations to confirm real-world gains for creators and inference workloads.