Open-weights video AI finally runs on local NVIDIA hardware

Meta has released Muse Glimmer, a new open-source generative image model designed to run on a single consumer computer rather than requiring large-scale cloud infrastructure. The lighter footprint makes the model more accessible to independent developers and researchers working outside of data center environments. It marks another step in Meta's ongoing push to distribute AI capabilities beyond proprietary, server-dependent systems.

Anthropic has announced plans to embed invisible watermarks into text and images produced by Claude, making it easier for platforms and users to identify AI-generated content. The move is tied to European regulatory requirements around AI transparency. The changes are not yet live, but represent a formal commitment from the company.

Apple appears to be working on a feature called Apple Reference Image that would embed provenance metadata directly into photos at the moment of capture on an iPhone. The system is designed to help users demonstrate that a photo is genuine and not AI-generated. Code references for the feature have surfaced in the iOS 27 beta 5, though it is not yet live for users.

Google has announced that its Gemini app has reached one billion users, marking a significant milestone for the company's AI assistant. Among those users, image generation has become a heavily used feature, with Gemini now producing more than 150 million images per day. Voice interaction is also prominent, with nearly two-thirds of users engaging the assistant through direct speech.

LTX-2.5 is a new open-weights video generation model designed to run on consumer NVIDIA hardware, producing clips up to 6.8 seconds long with native multishot support. Lightricks released it with day-one ComfyUI integration, making it accessible to hobbyists and professionals working locally. The release positions capable video generation as something achievable without cloud infrastructure.

Xiaomi's MiLM Plus research team has proposed PROVE, a new evaluation framework for video object removal that introduces two perception-aligned metrics - RC-S and RC-T - alongside a real-world benchmark dataset. The work addresses a growing mismatch between how well modern removal models perform and how poorly existing metrics capture that performance. Standard measures like PSNR, SSIM, and LPIPS regularly rank model outputs in ways that disagree with human judgment.