The pulse of generative image & video AI.

Mirage, a video world model developed by Microsoft Research and academic collaborators, introduces a persistent spatial memory system that stores scene information in latent space rather than relying on pixel-based point clouds. The approach keeps environments visually consistent across long camera movements while significantly reducing compute and memory costs. Moving object tracking across segments remains an open limitation.

A new model called "Count Anything" aims to be the first general-purpose AI system capable of counting objects in virtually any image using only a text prompt. Researchers report it cuts counting error rates roughly in half compared to prior approaches. The system handles a wide range of subjects - from crowd scenes to microscopic cell samples - though very dense arrangements and vague descriptions remain challenging.

Despite years of bold claims about AI transforming filmmaking, very few projects have emerged that feel like genuine entertainment audiences would seek out. A new short film, "Dear Upstairs Neighbors," offers a different approach - one built on custom-trained versions of Google DeepMind's Veo and Imagen models rather than off-the-shelf AI tools. It may point toward what a more serious production pipeline actually looks like.

MiniMax M3, a multimodal model supporting text, image, and video, is now available through NVIDIA's accelerated compute platform with a free API endpoint. The model incorporates sparse attention mechanisms, making it well-suited for long-context tasks. It marks a notable expansion of MiniMax's distribution reach via NVIDIA's infrastructure.

Avataar AI has released a distilled video generation model priced at $0.005 per second of output, positioning it as a cost-conscious option for the Indian market. The model is designed with cultural and regional awareness built in, targeting the practical demands of businesses operating at India's scale. It represents a focused approach to making AI video generation accessible beyond Western-centric platforms.

A growing number of platforms - including Artlist Studio, ComfyUI, Flora, and Amazon's Project Nara - are competing to become the go-to AI production environment for filmmakers. The race is less about which underlying model is most capable and more about which workflow best fits how film professionals actually work. IndieWire examines what these platforms are building and what the film industry is looking for.

Mixbook has introduced Story Mode, a prompt-based tool that lets users describe the style and feel of a photo book in plain language before the AI assembles it. The company claims the feature speeds up the creation process by 3.5 times and cuts the time from start to order in half.

Google DeepMind has released DiffusionGemma, an open model that applies diffusion-based generation to text, promising outputs up to four times faster than conventional autoregressive approaches. While diffusion has long been the dominant technique in image generation, its application to language models is still relatively new territory. The release adds a notable open option to a field that has so far seen limited competition.

Decart has released Oasis 3, a real-time world model designed to generate photorealistic driving environments for autonomous vehicle testing. The model is now accessible via API, opening it up to developers who want to build on top of its simulation capabilities. Extended sessions are possible, though the announcement comes with noted limitations.

Maket has introduced Auto-Complete, a feature that takes a rough sketch and any already-placed rooms and generates a complete residential floor plan from them. The tool then carries that layout into 3D modeling and rendered visuals. It is aimed at streamlining early-stage residential design work.

Apple's WWDC announcements were largely familiar ground - chatbots, text summarization, and image generation tools that already exist elsewhere. But buried in the first developer beta of iPadOS 26 is a feature that takes a more interesting direction, drawing comparisons to the "vibe coding" approach of describing what you want and letting AI figure out the rest.

Apple is updating Image Playground, its built-in AI image generator, with improvements that bring it closer to the quality users expect from competing tools. The changes address longstanding criticism that the feature produced stilted, cartoon-like results with limited practical use. It remains to be seen how the updated output holds up against third-party generators available on Apple devices.

Apple's upcoming iOS 27 will bring a new set of AI-powered photo editing tools to iPhone users. The update expands on the computational photography features Apple has been building into its operating system over recent releases. Details on the specific capabilities are emerging ahead of a formal announcement.

Microsoft Research's Lens is a 3.8-billion-parameter text-to-image model that matches much larger systems on standard benchmarks, trained at a fraction of the usual cost. The key difference is data quality: 800 million detailed captions written by GPT-4.1, rather than the noisy alt-text typically scraped from the web. Model weights and code are released under an open-source license.

Amazon has added an AI-powered design tool to its Shopping app, letting users describe an image concept to Alexa and apply the resulting artwork to printable merchandise. The feature covers products such as T-shirts, hoodies, and tumblers. It connects generative image capabilities directly to Amazon's existing print-on-demand infrastructure.

AI-generated social media personas have grown harder to distinguish from real people, raising questions about transparency and trust on platforms built around personal identity. Early virtual influencers were visually distinct enough that audiences could easily spot them, but that gap is closing fast. The Verge traces how the technology and the business models around it have matured together.

Meta's standalone AI app has introduced a 'For You' section that serves up AI-generated news-style articles, complete with AI-produced images and text. The content follows familiar clickbait patterns, raising questions about accuracy and the platform's direction. It marks a notable shift from the app's original focus on a social feed of user-shared AI conversations and images.

Google is closing Pixel Studio, its AI image generation app for Pixel devices, less than two years after it launched. The shutdown continues a pattern of Google retiring products that failed to gain lasting traction. Users will need to look elsewhere for on-device AI image tools.

AI-generated deepfakes of K-pop idols, including sexualized images and videos, have drawn pushback from within fan communities themselves. Fans are organizing to call out and report this content, framing it as a violation of the artists' dignity. The situation reflects a broader tension as generative tools become easier to access and misuse.

Google for Startups has published a new report examining how early-stage companies are approaching generative media tools and workflows. The findings draw on perspectives from founders and industry figures navigating this space. The report aims to offer practical context for startups integrating AI-generated image and video into their products.

Content labels on AI-generated images and videos have become more common across major platforms, but critics argue that labeling alone is not enough. The Verge makes the case that YouTube, Instagram, TikTok, and others should go a step further and give users the ability to actively filter AI-generated content from their feeds. Without that option, labels function more as a disclosure footnote than a meaningful tool for audience control.

xAI has updated its Grok Imagine system to version 1.5, adding an image-to-video model that converts still images into short video clips at up to 720p resolution. The new model accepts text prompts to guide motion and style, and multiple generated clips can be joined into longer sequences.

Google has introduced Dreambeans, a tool that pulls personal data from your Google account to generate AI-illustrated stories in a cartoon style. The feature represents a notable step toward using ambient personal data - photos, calendar events, and similar account content - as direct source material for generative image output. It is, by most measures, one of the more unusually named products Google has shipped.

A British MP has filed a lawsuit against xAI to establish whether the company bears legal responsibility for images generated by its Grok AI system. The case is part of a broader wave of scrutiny that includes investigations in the EU, the UK, and California. At issue is how far platform liability extends when an AI image generator produces harmful or problematic content.

Ideogram has released version 4.0 of its text-to-image model as an open-weight release, featuring native 2K resolution output, bounding box layout control, and refined text rendering. On the DesignArena leaderboard, it leads all open models, sitting just below closed systems from OpenAI and Google. Commercial use of the weights requires a paid license.

Blackmagic Design has shipped the final release of DaVinci Resolve 21, marking one of the most substantial updates the software has seen. The version adds a dedicated Photo page for still-image editing alongside a set of AI-powered tools spread across the editing, color, audio, and visual effects areas of the application.

Amazon is introducing AI-generated product images into its search results, displaying visuals that correspond to what shoppers type in rather than pulling exclusively from existing seller photography. The company says the feature is intended to help users navigate toward relevant products more easily.

Martin Scorsese has taken on an adviser role at Black Forest Labs, the AI image generation startup behind the Flux family of models. The move has drawn surprise and some criticism from filmmakers and artists who had expected the director to maintain a more cautious stance toward generative AI tools.

MiniMax has released M3, a multimodal model built around a new sparse attention architecture that supports context windows up to one million tokens. The model handles text, images, and video natively, and includes capabilities for agentic tasks such as computer use and coding. The release positions M3 as a general-purpose foundation model aimed at both long-context reasoning and real-world tool use.

Glass Imaging's GlassAI neural image signal processing technology has found a prominent home in the new Honor 600 smartphone, marking a significant step for the startup after years of development. The integration focuses on improving zoom photography, an area where smartphone cameras have historically struggled with noise and detail loss.

Nvidia used GTC Taipei to unveil several new tools aimed at physical AI applications, including a new world model, a larger autonomous driving model, and an open reference platform for humanoid robots. The announcements signal a continued push to make simulation and synthetic data central to how robots and vehicles are trained. Here is a closer look at what was shown and why it matters.

A model has filed a lawsuit against a fashion brand after the company used AI to generate images of her without permission. The case highlights growing legal and ethical tensions as brands increasingly turn to AI-generated visuals to cut production costs, often at the expense of the people whose likenesses or work previously fed those systems.

As AI-generated images grow indistinguishable from photographs, photography faces a deeper question about what authenticity actually means. Judges and seasoned experts are increasingly unable to tell the two apart, raising doubts about whether intent behind an image can serve as a meaningful dividing line. The crisis touches competitions, editorial standards, and how audiences relate to images.

Microsoft is preparing a set of new in-house AI models under the MAI brand for its Build 2026 developer conference. The lineup includes an image model, a transcription model, and a multilingual voice model. The announcements would mark a continued push by Microsoft to develop proprietary AI capabilities alongside its existing partnerships.

A short film produced by Darren Aronofsky and starring Chris Rock and Paul Rudd used Google's Veo video generation model to animate the layered glass sculptures of artist Dustin Yellin. The project, titled 'Goodnight Lamby,' premiered at Cannes in the Classics section. It represents one of the more high-profile uses of generative video in a festival-circuit film context.

Sellers on TikTok and other platforms are using AI-generated personas depicting Black women in emotional distress to market mass-produced dropshipped goods. The fake influencers are designed to exploit sympathy and identity, presenting factory items as handmade products from real small-business owners. The trend raises questions about platform enforcement and the harm done to actual Black creators.

Google has patched a billing bug in its Gemini app that caused a small number of Omni video generations to consume an entire usage quota. The fix also doubles the video generation allowance for Ultra subscribers and stops charging users for failed requests.

A startup called Shift is offering free home cleaning to New Yorkers - with a condition attached. The company films its cleaners working through ordinary household tasks, and that footage becomes training data for robotic systems. It is one of several efforts underway to capture the kind of mundane domestic labor that has proved difficult to teach machines.

Director Jorge R. Gutierrez has walked back plans to produce an AI-generated animated series called 'Punky Duck' for Amazon MGM following significant backlash from the animation community. Gutierrez said his original intent was to highlight artists using the technology, but the response made clear the project would not move forward in its planned form.

Google Cloud has moved Nano Banana 2 and Nano Banana Pro into general availability, bringing scalable image generation with 1K and 2K output resolution to production workloads. The models also include a preview feature for video input, expanding their range of supported media types.

Adobe's Firefly AI Assistant takes a different approach than most generative image tools - acting as a conversational agent that operates Adobe's design apps on your behalf rather than generating images from scratch. A hands-on beta test finds the concept genuinely interesting but the actual output underwhelming, somewhere between useful shortcut and frustrating limitation.

A 75-minute AI-generated film about the Iranian government's killing of protesters will premiere at the Tribeca Festival next month. Called Dreams of Violets, it was made for roughly $2,000 by two Iranian-born brothers using AI to recreate people and events drawn from journalistic records and eyewitness accounts.

Amazon MGM Studios and AWS are launching a "GenAI Creators' Fund" that gives filmmakers money and access to the in-house AI platform "Project Nara." Three animated series are already in production - the teams had five weeks for their pilots. Amazon says it now has the "only end-to-end AI content ecosystem in the industry." The article Amazon builds its own AI production platform and greenlights three AI animated series for Prime Video appeared first on The Decoder.

DiffusionBlocks converts residual networks into independently trainable blocks by interpreting layer updates as reverse diffusion denoising steps. The post Sakana AI Proposes DiffusionBlocks: a Block-wise Training Framework That Converts Residual Networks into Independently Trainable Denoising Modules appeared first on MarkTechPost.

Google's new Gemini Omni artificial intelligence (AI) model can do some wild things. The model's key promise is to create anything from, well, anything. [Read More]

Microsoft's MAI-Image-2.5 ranks third on Arena's text-to-image leaderboard, on par with Google's Nano Banana 2 but still behind OpenAI's Image-2. The model shows clear gains over its predecessor, especially in rendering text inside images and commercial visuals. The article Microsoft's MAI-Image-2.5 pulls even with Google's Nano Banana 2 on benchmarks appeared first on The Decoder.

YouTube will now automatically label videos that use significant photorealistic AI, instead of relying solely on creators to disclose AI-generated content themselves. It's also making AI labels more prominent.

Digital asset platform Artlist is launching Artlist TV, a streaming platform that appears to be exclusively populated by AI-generated shows. [Read More]