Researchers just introduced TARFlow, a scalable architecture that pairs autoregressive Transformers with exact-likelihood normalizing flows. By revamping traditional coupling layers and borrowing guidance techniques, it achieves diffusion-level quality across images and video while remaining fully invertible.
xAI just rolled out Aurora, a new autoregressive image generation model for Grok on X. Trained on billions of web examples, it tackles notoriously tough challenges like text rendering, logos, and photorealistic portraits, alongside native support for image editing inputs. It is now live in select countries and rolling out to all users within a week.
AMD just released and open-sourced Nitro Diffusion, a pair of one-step text-to-image models based on Stable Diffusion 2.1 and PixArt-Sigma. Trained on AMD Instinct MI250 accelerators, they match full-step visual quality in a single inference step for both data centers and edge PCs. Code and weights are now up on Hugging Face and GitHub.
Google just launched PaliGemma 2, its next-gen vision-language model family built on Gemma 2. It offers drop-in replacement upgrades across multiple model sizes without major code changes, making it significantly easier for developers to fine-tune visual AI for custom datasets and vision tasks.
OpenAI just hired three senior computer vision researchers from Google DeepMind—Lucas Beyer, Alexander Kolesnikov, and Xiaohua Zhai—to staff its newly opened office in Zurich, Switzerland. The team will focus on developing multimodal AI models across images and audio. Why it matters: The high-profile raid underscores an escalating global talent war, where top AI labs routinely shell out seven-figure compensation packages to secure a limited pool of frontier researchers.
Google just opened its flagship Imagen 3 generator to all Vertex AI customers, while launching video model Veo in private preview. Veo produces six-second clips from text or image prompts with localized editing support. Google is finally pushing its top creative models to enterprise clients, though Veo remains gated behind early access.
Amazon just unveiled Amazon Nova, its new lineup of foundation models spanning text, multimodal reasoning, and image/video generation. Available in Amazon Bedrock, the Micro, Lite, and Pro models claim to be the fastest in their classes while slashing costs by at least 75%.
Fei-Fei Li’s startup World Labs just revealed its first project: an AI system that generates interactive, explorable 3D scenes from a single 2D photo. Users can navigate browser-rendered spaces with mouse and keyboard controls, complete with adjustable depth of field. It proves world models can maintain spatial consistency, though movement boundaries remain limited.
Google Labs just released GenChess, an experimental project powered by the Imagen 3 image generation model. Players can type text prompts to design custom classic or creative chess pieces, then play a match directly against the computer across several difficulty levels. It is currently available in select regions.
Stability AI just released three new ControlNet models for Stable Diffusion 3.5 Large: Blur, Canny, and Depth. These models allow creators to handle upscaling up to 16K, line art structure, and depth map guidance with high precision. They are free for non-commercial use and commercial use under $1M in annual revenue, supported by Comfy UI.
MIT researchers just unveiled the Earth Intelligence Engine, a tool combining generative AI with physics-based flood modeling to generate realistic satellite imagery of future flood scenarios. While standard GANs hallucinated water on dry hilltops, fusing them with physical simulations keeps the disaster projections accurate enough to guide actual evacuation choices.
Luma AI just expanded its video generator Dream Machine into a full creative platform, introducing a new foundation image model called Photon. The model is reportedly up to 800% faster than comparable tools, arriving alongside a refreshed interface, an iOS mobile app, and a dedicated image API for developers.
AMD just launched Amuse 2.2 Beta, bringing support for Stable Diffusion 3.5 to Ryzen AI 300 and Radeon 7000 hardware. The update introduces offline high-quality modes and XDNA super-resolution, boosting native output resolutions up to 2048x2048 pixels. Why it matters: Local generative AI on everyday AMD laptops is finally getting fast, painless, and sharp.
Runway just launched Frames, its newest base image model designed to tackle AI's consistency problem. It lets creators lock in a specific aesthetic and reliably generate cohesive variations across different scenes, from vintage 35mm photos to 70s anime. Access is gradually rolling out inside Gen-3 Alpha and via the Runway API.
Krea just rolled out Flux image prompts, allowing creators to condition their image generations using reference pictures. The added image-to-image steering is completely free for everyone to use.
Meta just announced major upgrades to Messenger Calling, rolling out AI-generated backgrounds powered by Meta AI alongside HD video and voice isolation. Users can also leave audio or video messages when friends miss a call, turning the chat app into a much more capable communication hub.
Niantic just announced it is building a Large Geospatial Model trained on millions of real-world 3D scans uploaded by Pokémon Go and Ingress players. The goal is to give robots and computers human-like spatial intelligence, using pedestrian-level maps of places cars cannot reach.
Mistral AI just rolled out a major beta update to le Chat, packing web search with citations, a collaborative Canvas workspace, and Flux Pro image generation into its free assistant. Powered by the new Pixtral Large multimodal model for complex document and image reasoning, it also lets users build workflow agents—giving ChatGPT a serious free run for its money.
Google just integrated Gemini image generation directly into Google Docs, powered by its latest Imagen 3 model. Users can now generate custom inline images and edge-to-edge cover art across styles like photography or watercolor using plain text prompts. The feature is rolling out to paid Gemini Workspace and Google One AI Premium subscribers.
Basejump just revealed a Web3 social gaming platform powered by Action, an AI substrate that lets users generate game worlds and assets without coding. Backed by fresh funding from Community Labs and running on AO's distributed compute, the platform is scheduled to debut in early 2025.
