Two regulators got specific about your output the same week your tools got quicker.
California's AI Transparency Act (SB 942) took effect yesterday. Large AI developers must embed both visible and machine-readable disclosure in AI-generated images, video, and audio. Systems with over 1 million monthly visitors are covered. Providers must also enforce compliance from third-party licensees within 96 hours. If you ship AI media tools, this changes your labeling requirements.
Article 50 of the EU AI Act has entered into force, requiring any synthetic audio, image, video, or text to carry machine-readable labels indicating AI generation. Deployers must also disclose deepfakes and AI-generated public-interest text unless human-reviewed. There is no grace period for compliance.
Google launched an AI image generation tool in Google Earth that let users overlay AI generated images onto real maps. It was pulled within 24 hours after journalists and researchers flagged the risk of spreading misinformation through a trusted mapping source. Google cited policy violations and said it will add stronger guardrails before bringing the feature back.
A hands-on run of Meshy AI, the browser-based text-to-3D tool, on a stylized knight, a treasure chest and a dinosaur. It generates PBR materials, adjusts topology on remesh, and auto-rigs character assets against animation presets. The failure mode is thin detail: mouths, teeth and claws look fine in a rendered preview while hiding geometry that is hard to animate or print. Inspect those before anything reaches production.
A practical head-to-head test of character consistency methods for AI video generation shows the 360 video trick and omni models deliver the most reliable results. Blacked-out face references only help in specific multi-character shots. The comparison uses real credit costs, giving builders actionable workflow data without marketing claims.
Pippa pays artists $0.003 per second of AI-generated video and $0.005 per image, plus 5% of subscription revenue. Only 8 artists have signed up, and the service has 800 subscribers. The model addresses compensation but leaves consent, credit, and market substitution unresolved. The base model still trains on scraped data, and artists can opt for pseudonyms, signaling social stigma around participation.
A ComfyUI node that forecasts post-transformer features so it can skip transformer blocks on selected solver steps. The repo measures a 35 percent cut in transformer-step count on 20-step Euler runs and 30 percent on RES multistep. The one community wall-clock figure is 24 percent. It is explicitly not bit-identical, so a locked seed no longer reproduces a locked shot: close to free while you iterate, costly once you are finalizing.
Mirelo MCP is a hosted MCP server that adds sound effect generation to Claude, ChatGPT, and Cursor. Users describe sounds in text or provide video, and the server generates new audio, extends clips, or inpaints specific regions. Setup requires adding a URL and signing in with a Mirelo account, with usage billed to Studio credits. For video editors, game developers, and app builders without audio expertise, this puts custom sound design into their existing AI workflow.