OpenAI launched the Images 2.6 model, enabling ChatGPT to natively generate and edit GIF animations from prompts or uploaded images for the first time. The update also improves generation speed and introduces GPT-Image-2.6 Flare and Sunburst variants for users and API developers.
Multimodal Generation
多模态生成
The ability of models to process multiple data types like text and images simultaneously and generate corresponding content.
Curated Intelligence
SenseTime released the multimodal agent model SenseNova 6.8 Flash, focusing on 'delegated intelligence' and end-to-end delivery. The model excels in web generation, 3D PPTs, and dynamic posters, aiming to lower barriers for design and office automation by allowing users to provide goals while AI handles the full workflow.
Genspark has officially launched Media Studio, a video agent workbench that orchestrates LLMs and multimodal models to automate short drama production. The tool supports one-click template replication and interactive storyboarding, significantly lowering the barrier for non-professional creators. A promotional discount on member credits is currently active.
Google has released Genie 4, a new world model that integrates Gemini Omni's visual anchoring and reasoning capabilities to enable unprecedented real-time interactive simulation. Natively multimodal, it simultaneously generates ambient audio, music, and dialogue, significantly enhancing physical consistency and behavioral coherence in virtual environments as a key step toward AGI.
NVIDIA released the open-source model RFD4-Proteina on Hugging Face as part of the BioNeMo series. It addresses information loss in traditional protein design by unifying multi-modal generation capabilities into a single end-to-end workflow. The model supports defining design goals and generating candidate proteins, and is integrated with the Proteinbase competition platform to accelerate experimental validation.
Recent research reveals that viewing AI-edited images or reading inaccurate AI summaries significantly increases the likelihood of false memories, with subjects exhibiting higher confidence in these fabricated details. The study demonstrates that even when users are warned about AI involvement, they remain susceptible to misinformation, suggesting that human oversight as a safeguard against AI errors may be fundamentally compromised.
Grimoire for Biology launched an interactive learning tool powered by LLMs that generates dynamic knowledge graphs, 3D models, and video explanations from natural language queries. It visualizes complex concepts like protein engineering and cell structures in a draggable interface to lower entry barriers and enhance biology education.
TypeSafe AI released Jev 1.14.0, introducing experimental features such as image generation, new data types, and extended context windows to balance reliability with creativity for 'dreamer' users.
Kombai introduces Design Creativity Bench, evaluating AI creativity in UI design across originality, creative range, and appropriateness. Results reveal a significant gap between frontier models and human designers, with the top model GPT-6 Astra scoring 65.5 against a human reference of 92.3. The benchmark's methodology and datasets are open-sourced to make design creativity a measurable dimension of model capability.
OpenAI announced a partnership with video platform Higgsfield, leveraging the new GPT-6 Astra model to extend AI video generation limits from 1 minute to 10 minutes. This breakthrough relies on GPT-6 Astra's 'creative agent' capabilities, which utilize external tools like 3D modeling to maintain consistency in characters, scenes, and visual details over longer durations. Additionally, Higgsfield integrated GPT-Image-2.5 for stable text rendering in images and used Codex to accelerate internal product development.