Topaz's March Upgrades Speed Up Pipelines, Add Local Model Rendering for Creators
Topaz’s Wonder 2 now runs locally via NeuroStream, letting creators denoise, sharpen, and upscale in one step on RTX and RTX PRO GPUs while Astra gains per-segment Scene Controls and Batch Render.

Announced in January 2026 and now available in Topaz Photo, Wonder 2 is billed as an all-in-one image enhancement model that "denoises, sharpens, and upscales in a single step, with no sliders or tuning required." The company says Wonder 2 (Local) runs on standard creator hardware thanks to its NeuroStream VRAM optimizations, a move that shifts heavy image and video inference from cloud-only workflows to local machines.
NeuroStream is described as an industry-first VRAM optimization that "dramatically reduces VRAM usage and allows powerful AI to run on standard creator hardware." Topaz engineers worked with NVIDIA to tune the implementation, and Gerardo Delgado Cabrera, director of product for AI PCs at NVIDIA, framed the collaboration succinctly: "As the demand for local processing on RTX GPUs continues to grow, NeuroStream provides an opportunity to run complex AI models on nearly all hardware," said Gerardo Delgado Cabrera. "This latest collaboration with Topaz Labs is part of ongoing efforts to help develop technology optimized for use with NVIDIA-powered devices." Topaz says the combination lets large models that previously required beefy workstations run on every NVIDIA GeForce RTX and RTX PRO GPU.
Video creators see parallel upgrades in Astra, where new Scene Controls enable per-segment tweaks and Batch Render processes multiple files at once to speed creative pipelines. Topaz released new Starlight variants tied to Astra, although naming varies across materials: one update references Starlight Fast 2 while another describes Starlight Precise 2, the latter noted to bring "softness and natural motion, though it may soften some textures." That tradeoff underscores a common tension in AI video work between motion realism and texture retention.
Topaz framed these cumulative changes as part of a broader Realism Update rolled out in late 2025 that added more than 20 models, features, and improvements across Bloom, Topaz Photo AI, Topaz Gigapixel, Topaz Video AI, and Astra. The Realism Update aims to "address creator pain points - like AI's 'plastic' look - while expanding accessibility." Enterprise-focused API expansions now include Image Colorization, Image Object Matting, Transparent Image Upscale, Video Colorization, Starlight HQ, and Face Realism, with potential for ComfyUI nodes via API for custom developer setups.
Incremental Photo AI releases layer performance and workflow gains on top of the realism push. Photo AI 3.5 introduced "Smarter Lighting Adjustments," claims of "Processing speed is now 40% faster on M1 and 10% faster on M4" for Super Focus, and a cloud Super Focus mode that "sharpens large images 7x faster than local processing." A later item lists Photo AI v3.6.0 as able to "refocus blurry photos 500% faster." Community release notes from Photo AI 2.4 in March 2024 show prior workflow changes, including an overhauled crop tool and a revamped Remove tool; users pushed back in places, noting oversharpening on complex textures such as bird plumage.
Smaller but practical tweaks include Adjust Lighting 3, Dust & Scratch 2 for preserving detail, a Grain Toggle to "eliminate plastic smoothness and add authentic grain," restored full Intel Mac support in Topaz Gigapixel, and plugin reliability updates for After Effects and DaVinci Resolve. Taken together, the package pushes a local-first, performance-oriented agenda while expanding enterprise APIs and video workflow controls. Key verification points remain around exact Starlight variant names, full NeuroStream GPU requirements, and official changelog dates for Photo AI 3.5 and 3.6.0, but the direction is clear: Topaz is moving heavy AI work onto creators' machines and into faster, more granular pipelines.
Know something we missed? Have a correction or additional information?
Submit a Tip

