The pulse of generative image & video AI.

Hany Farid, one of the foremost researchers on synthetic media and digital forensics, sits down with 404 Media to discuss the trajectory of deepfakes - from their origins to where the technology is headed. The conversation covers how detection, policy, and public perception have struggled to keep pace with the rapid advancement of generative tools. It is a grounded look at what the erosion of verifiable reality means for individuals and institutions alike.

Upload a photo, pick a trending dance template, and Groovely animates it into a vertical clip synced to the beat. No camera, choreography, or editing required.

What began as a wave of skepticism after OpenAI's Sora failed to immediately reshape filmmaking has given way to a maturing AI video industry with real commercial footholds. Companies like Promise are embedding themselves near Hollywood studios, Netflix is integrating AI tools across hundreds of titles, and startups in the space are commanding multi-billion-dollar valuations. The business infrastructure - job roles, revenue splits, and specialized vendors - is now catching up to the technology i

A new benchmark from Moonshot AI isolates visual perception from logical reasoning in multimodal models, and the results are sobering. No tested frontier model clears 60 percent accuracy, with GPT-4o leading by only a narrow margin. The findings suggest that many errors previously attributed to faulty reasoning may actually originate much earlier, at the point of reading the image itself.

Google is updating its AI-generated content policy to let users remove visible watermarks from images and other media produced by its tools. The invisible SynthID watermark, however, will stay embedded in the content regardless. This marks a shift in how Google balances user flexibility with its underlying approach to AI provenance tracking.

Google has added a toggle in Gemini and its AI video tool Flow that lets users remove the visible "sparkle" watermark from AI-generated images, videos, and music. Even with the visible mark turned off, content will still carry invisible SynthID watermarks and C2PA metadata. The change affects content produced by Google's Nano Banana and Omni models.

Google is giving users the option to remove the visible watermark that appears on AI-generated images and videos, responding to feedback from creators who found the branding intrusive. The change applies only to the visible mark - the invisible SynthID watermark, used to identify content as AI-generated, remains embedded regardless of the setting. It is a notable shift in how Google balances transparency with creative flexibility.

Cara, a social platform built specifically to protect photographers from AI data scraping, has reportedly had its entire image library harvested by an outside actor. The incident is a pointed reminder that technical and policy-level protections do not always hold against determined scrapers. The person responsible apparently made their actions public, drawing sharp criticism from the creative community.

A new short film produced with Higgsfield's AI video tools offers a rare look at what AI-assisted filmmaking actually looks and feels like in practice. The Verge went inside the production to find out where the technology helped - and where human craft still carried the weight. The result is a candid assessment of where generative video sits today.

AI-powered upscaling tools have made it easier than ever to improve the resolution and clarity of older or low-quality photos. Engadget walks through some of the most accessible methods available today, from dedicated apps to built-in software features. Whether you are working with a scanned print or a compressed digital file, there are practical options that do not require professional editing skills.

The Roku Channel has launched Fairground AI, a 24/7 linear channel running entirely on AI-generated content, including narrative shorts and commercials. The move raises real questions about audience appetite in Western markets, even as China's short-form AI video ecosystem already commands billions of views. Whether this is a glimpse of where streaming is headed, or a curiosity that quietly disappears, remains an open question.

Twitch has introduced an opt-out setting that lets streamers prevent their content - including streams, VODs, clips, chat logs, and channel images - from being used to train Amazon's generative AI models. The control applies to future training only and covers AI systems designed to generate or synthesize text, audio, images, or video. Non-generative AI features such as captions and safety tools are unaffected by the setting.

Insta360 has announced the X6, its latest 360-degree action camera, featuring an 8K50p video recording capability built around a 1/1.1" sensor and a triple AI chip system. The camera also supports 6K60p, a Bullet Time mode at 8K100p, and 16K time-lapses, placing it at the higher end of the consumer 360-degree camera market. Pricing and full availability details have been released alongside the announcement.

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.

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.

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.

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.

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.

A new technical guide walks through building a fully programmable MiniMax-H3 pipeline using ComfyUI as a headless backend, covering everything from hardware profiling to joint video-audio decoding. The tutorial treats ComfyUI not as a visual tool but as an API-driven inference engine for multimodal generation. It offers a practical path for developers who want reproducible, automated workflows without relying on a graphical interface.

The Verge takes a critical look at Mark Zuckerberg's vision for AI and what it reveals about a broader cultural trend - the outsourcing of meaning, motivation, and creativity to generative tools. Using a telling anecdote about an AI-generated motivational poster, the piece questions what is lost when personal expression is handed off to a machine. It is a thoughtful provocation aimed at the assumptions quietly baked into how AI is being sold to us.

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.

Meta has released Muse Glimmer, an open-source multimodal AI system designed to run locally and operate with agentic capabilities. The model combines image understanding and generation with the ability to take multi-step actions, all without requiring a cloud connection. It marks Meta's latest push into accessible, on-device generative AI tools.

xAI has released Imagine Image 2.0, its latest image generation model for the Grok platform, and it has already secured second place in Arena benchmarks - sitting just below OpenAI's GPT-Image-2. The update also brings new editing capabilities and template options aimed at practical creative use.

xAI has updated its image generation tool inside Grok with the release of Imagine Image 2.0, available through the app's Quality Mode on both web and mobile. The update introduces a range of new capabilities including precise editing, smart resizing, and multi-reference image input. Workflow templates are also part of the release, aimed at streamlining repeated or complex image tasks.

Roku has launched a 24/7 free ad-supported streaming channel dedicated entirely to AI-generated content, sourced from a startup called Fairground. The move marks one of the more visible attempts to bring generative video into mainstream living-room viewing. Whether audiences will warm to it is another question.

Adobe has expanded its presence inside ChatGPT, moving beyond its initial Photoshop, Express, and Acrobat integrations to bring its entire suite of applications into OpenAI's conversational platform. The move positions Adobe more directly against Canva, which has built much of its recent growth on accessible, AI-assisted design tools. The unified plugin signals a deeper strategic bet on ChatGPT as a creative interface.

Google's Gemini Omni lets users generate and edit video through natural conversation, and a new spotlight from the company shows how five independent builders are putting that capability to practical use. The examples range from visualizing abstract ideas to streamlining video editing workflows. Together, they offer a concrete look at how conversational video AI fits into real creative and production work.

Adobe and OpenAI have deepened their existing partnership with a unified plugin that brings more than 70 Adobe tools directly into ChatGPT. The integration spans apps including Photoshop, Firefly, Premiere, Illustrator, and Acrobat, letting users describe a creative goal and have the AI orchestrate the appropriate tools to complete it. The plugin is available now for ChatGPT users, with guest access offered and fuller capabilities unlocked via an Adobe account sign-in.

Skylum, the company behind Luminar, is adjusting how it talks about its AI-powered photo editing tools as public sentiment toward artificial intelligence grows more skeptical. Rather than leaning into AI as a selling point, the company is reframing its software around the idea of supporting human creativity. It is a notable shift for a brand that built much of its identity on being an AI-first editing platform.

Black Forest Labs has moved FLUX 3 Video out of early access and into general availability, offering Full HD video generation with clips up to 20 seconds long. The model includes native audio output and lip-synced dialogue across more than 14 languages. According to BFL's own Elo benchmark rankings, it outperforms both Gemini Omni Flash and Seedance 2.0.

A car photography company has filed a lawsuit against Midjourney, alleging the AI image generator used tens of thousands of its copyrighted vehicle photos as training data without authorization. The case adds to a growing body of copyright litigation targeting generative AI companies over their use of unlicensed visual material. It is one of the more focused suits in the space, centering on a single industry's photographic catalog rather than a broad coalition of creators.

ByteDance has released SeedRealtime, a full-duplex multimodal model that handles audio, video, and text within a single unified system. Unlike turn-based conversational AI, it supports continuous, overlapping dialogue with natural timing and proactive responses. The model represents a step toward more fluid human-computer interaction without the stop-and-start feel of earlier voice interfaces.

Mistral has released Shieldstral, a 3-billion-parameter open-weights multimodal safety classifier designed to moderate text, images, and combined text-image content. The model is built for customizability, giving developers and organizations direct control over how moderation policies are applied. Its open-weights nature means it can be deployed and adapted without relying on a hosted API.

A federal judge has declined xAI's request for a preliminary injunction to stop a Minnesota law banning AI-powered 'nudify' applications from taking effect. The ruling allows the state's restrictions on non-consensual AI-generated intimate imagery to stand while litigation continues.

Moonshot's PerceptionBench offers a structured way to measure how well multimodal vision models handle fine-grained visual tasks, from OCR and object counting to depth understanding and hallucination detection. A new tutorial walks through building a complete evaluation pipeline, including environment setup, dataset loading, and automated answer judging. The workflow is designed to run in Google Colab, making it accessible for researchers without dedicated infrastructure.

A federal judge has denied xAI's request for a preliminary injunction against a Minnesota law banning apps that generate non-consensual intimate imagery. The ruling allows the law to take effect while xAI's broader legal challenge continues.

A Belfast photography gallery has drawn criticism from photographers and the wider arts community after hosting an exhibition that included AI-generated images alongside traditional photographic work. The decision has reignited longstanding debates about whether AI-generated imagery belongs in spaces dedicated to photography. Critics argue that exhibiting such work under a photography banner misrepresents the medium and devalues the craft of working photographers.

Google Earth quietly removed its newly integrated Nano Banana AI image generator just days after launch, following reports that users were producing misleading images with the tool. The feature had only just been covered in the press when Google moved to pull it, suggesting the decision was a rapid response to early misuse.

Chinese AI company MiniMax has released the weights for its H3 video generation model, marking the first time an open model has claimed the top spot on a major AI video benchmark ranking. The release is a notable moment for the open-source side of the generative video space, which has largely been outpaced by proprietary offerings from companies like OpenAI, Google, and Sora competitors. H3's rise to the top of the leaderboard signals that the gap between closed and open video models may be narr

Alibaba's Qwen team has moved Qwen3.8-Max from preview to general availability, marking it as the most capable model in the Qwen lineup to date. The mixture-of-experts model carries 2.4 trillion parameters and supports text, image, and video input over a 1 million-token context window. Open weights are expected to follow next week.

A new tutorial walks through a complete GeoAI pipeline for extracting building footprints from high-resolution NAIP aerial imagery, combining classical deep learning with newer vision models. The workflow covers everything from environment setup and data preparation to training and inference across four distinct model architectures. It offers a practical reference for anyone working at the intersection of geospatial analysis and generative or segmentation-based AI.

A new wave of AI startups is attempting to address longstanding concerns from the illustration community by compensating artists whose work is used in model training. Pippa is one such company, positioning itself as a more ethically grounded alternative to competitors that have trained on unlicensed work. Whether financial compensation alone is enough to shift artist sentiment remains an open question.

Snap and LinkedIn are both taking steps to curb the spread of low-effort AI-generated content on their platforms. Snap is blocking AI-generated videos from its Spotlight feed, while LinkedIn has introduced a dedicated reporting option for what many users call "AI slop." The moves reflect growing pressure on social platforms to maintain content quality as generative AI tools become more accessible.

A federal judge has denied xAI's request to block a Minnesota law targeting so-called "nudify" apps, allowing the state's ban to remain in effect. The ruling is a setback for xAI, which had challenged the law on legal grounds. The case adds to a growing body of litigation around state-level regulation of generative AI tools.

Fenix Flexin, known as a member of Los Angeles rap duo Shoreline Mafia, has a solo track climbing the Billboard Hot 100 - but listeners and critics are questioning whether "Rubberz" was generated by AI. The song marks a sharp stylistic departure from his usual trap-influenced West Coast sound, fueling speculation. Fenix has denied the claims while doing little to address the specific concerns raised.

A federal judge has declined xAI's request to block a Minnesota law targeting so-called "nudify" apps, which use AI to generate non-consensual nude imagery. The ruling allows the law to remain in effect while the broader legal challenge proceeds. xAI had filed the lawsuit just days before seeking the emergency injunction.

ByteDance has released Seedance 2.5, an AI video model capable of generating clips up to 30 seconds long with audio produced in the same pass. The model accepts a wide range of reference inputs, including images, videos, and audio files. For production workflows that currently involve stitching together many short segments, this could meaningfully reduce the steps involved.

xAI has updated its Imagine Video model to version 1.5, adding support for character and voice references, prompt-only generation, and native 1080p output. The update gives users more control over how characters look and sound across generated clips. Multi-reference inputs are also now supported, allowing several reference elements to be combined in a single generation.