Archive, page 4

Krea has released Krea 3, an image editing model focused on speed and precision, delivering 2K-resolution edits in under 8 seconds. The model supports training on specific products, people, or styles, and is priced at $0.03 per image. It represents a step toward making high-quality, customized image editing more accessible and cost-effective.

AI-generated images mimicking celebrities like Mia Goth and Taylor Swift are spreading across Instagram, sometimes through verified accounts with millions of followers. These deepfake posts exploit real-life events to appear credible, blurring the line between authentic celebrity content and fabricated material. The trend raises serious questions about platform accountability and the practical limits of content moderation at scale.

TIME magazine has recognized Google's SynthID as one of its Best Inventions of 2026, highlighting the tool's approach to watermarking and verifying AI-generated content. SynthID embeds imperceptible signals into images, audio, and other media to help identify what was made by an AI system. The recognition brings wider attention to the growing field of AI content provenance.

Meta has released Muse Spark Open, an open-weight multimodal mixture-of-experts model aimed at the frontier tier of generative AI. The release makes the model's weights publicly available, continuing Meta's pattern of open releases in a space often dominated by closed systems. It represents a notable step in bringing high-capacity multimodal architectures to the broader research and developer community.

Hugging Face has integrated FLUX.3 Image Klein into its Diffusers library, bringing Black Forest Labs' lightweight FLUX model to a widely used open-source framework. The addition makes it easier for developers and researchers to run and experiment with the model using familiar tooling. This post walks through what FLUX.3 Klein is and how to get started with it in Diffusers.

A growing online trend dubbed "grandmaslop" refers to the uncanny, often unsettling AI-generated images of children and family members that older adults share on social media. Younger relatives have taken notice - and many find the results more disturbing than heartwarming. The phenomenon sheds light on how generative image tools are being adopted by demographics with little context for their quirks and limitations.

Shootsolo, a voice-controlled camera app aimed at solo creators, has released its 2.0 update with the ability to record portrait and landscape video simultaneously from a single take. The update also adds a built-in teleprompter, virtual studio lighting, and background effects. The app is free to download on both iOS and Android, with no subscription, watermarks, or ads.

The PetaPixel Podcast team - Chris, Jordan, Sarah, and the host - gather to break down the week's biggest stories in photography and imaging. Leading the discussion is Adobe Creative Cloud's apparent AI-driven overhaul, alongside Sony's latest microphone releases and a wave of APS-C camera news. It's a broad, news-heavy episode covering several corners of the imaging world.

Researchers have found that recent image generation models from OpenAI and Google can be manipulated into producing explicit imagery despite built-in content safeguards. The findings raise fresh concerns about the durability of safety measures in widely deployed generative AI systems. Both companies have been expanding access to their latest models amid growing scrutiny over AI-generated non-consensual content.

An AI-generated image of actress Mia Goth spread widely on Instagram after her divorce filing made headlines, with hundreds of thousands of users - including industry figures - liking the post before it was identified as fake. Goth has spoken out against the platform for allowing the image to circulate. The incident raises fresh questions about how social media platforms handle AI-generated content of real people.

Epic Games has issued an "AI-first" mandate for an unannounced Disney title, directing developers to use tools like Anthropic's Claude for code generation and to produce all in-game art through AI workflows. The project has slipped past its original March release window as the studio reorients around these new processes. The directive reportedly includes an explicit instruction from a director that there will be "no more hand-written code."

Higgsfield has extended its AI video generation limit from one minute to over ten, a significant step for creators working on longer-form content. The update also brings tighter visual consistency powered by GPT-Image-2.5 and introduces voice-driven filmmaking through GPT-Live. Together, the changes position Higgsfield as a more capable tool for end-to-end video production.

Code references to a model called GemPix 3 Pro have surfaced, pointing to Google working on a new Gemini 4 Argon-based image generation model. The upgrade appears destined to power Nano Banana 3 Pro, Google's next iteration in its on-device AI lineup. Details remain sparse, but the findings offer an early look at Google's image AI roadmap.

VideoProc offers a consolidated workflow for restoring and upscaling aging or low-quality photos and video footage. The tool targets common problems like noise, low resolution, and insufficient detail for large-format output. It positions itself as a single-application solution rather than a patchwork of separate editing tools.

ARRI has released version 1.5 of its Film Lab OpenFX plugin, a substantial update that expands the tool's analog film simulation beyond its original 35mm base to include 8mm, 16mm, and 65mm formats. The update also brings faster processing, new grain controls, and anamorphic squeeze options. Existing license holders can upgrade at no cost through RE:Vision Effects.

As AI image generators make photorealistic fabrications increasingly easy to produce, the question of how to verify that a photograph reflects something that actually happened has become more urgent. Apple is among the players now engaging with content authenticity efforts aimed at answering that question. This piece from PetaPixel explores the tools and standards being developed to establish trust in visual media.

Cloudflare has released Clef-omni, a multimodal content moderation and decision model capable of processing text, images, audio, and video within a single API call. Built on the Qwen3-Omni-30B-A3B architecture and released under the Apache 2.0 license, it is priced at $0.15 per million input tokens. A lighter variant, Clef-flash, is also available at $0.038 per million tokens.

vLLM Semantic Router has released Decision 3.0, a family of five open-weight multimodal models ranging from 0.8B to 27B parameters, all licensed under Apache 2.0. The models are designed to answer structured decision-type questions - multiple choice, yes/no, and scoring - about both text and images in a single inference call, without generating free-form text.

Alibaba's Qwen team has released Qwen-Image-2.1-Turbo, an accelerated variant of its open-weight image model that generates and edits images in just 8 denoising steps, down from the 40-step default of the base model. The 7B-parameter architecture remains unchanged, meaning the speed gains come from a distilled inference process rather than a smaller model. A hosted API option accompanies the open-weight release.

Nikon has disqualified the original first-place winner of its Small World in Motion microscopy video competition after determining the entry used generative AI in post-processing, violating contest rules. Dr. Ning Xu's video, which purported to show cilia moving in a child's airway, came under scrutiny after doubts about its authenticity surfaced online. Nikon confirmed last week that it had been reviewing the submission before reaching its decision.

Midjourney is testing a new "Thinking Mode" for image generation, currently available on its Alpha website. The feature can be accessed by reopening a completed job and selecting "Rerun (Thinking)" from the lower right corner of the lightbox. The company is actively seeking user feedback on the experimental capability.

Midjourney's latest alpha update introduces shared folders, giving users a way to collaborate directly within the platform. The changelog, drawn partly from community feedback via the #alpha-ideas-and-bugs channel and office hours, also includes a broader set of smaller fixes and improvements. It marks a notable step toward making Midjourney a more team-friendly workspace.

Runway has announced an integration with Claude Motion, allowing users to export animated explainer content directly into Runway's video production environment. From there, creators can generate footage, layer in voiceover and music, edit at the scene level, and reformat outputs for different platforms.

Anthropic has introduced two beta features for Claude: Dashboards, which connects to data sources like BigQuery and Snowflake to build live dashboards from text prompts, and Motion, which produces animated explainer videos from text and images. The company also expanded its existing Docs, Slides, and Design tools to all users, including those on free plans.