Search AI Products and News

Black Forest Labs & Krea launch open-source image model FLUX.1-Krea, addressing AI's artificial look via 12B-parameter diffusion. Enhanced training improves realism in lighting/colors, compatible with FLUX ecosystem for commercial use.....

The StepZen team announced that their latest generation foundational large model, Step3, is officially open-sourced. Step3 is a model designed for enterprises and developers who pursue the optimal balance between performance and cost, aiming to create the most suitable model for the inference era. The open-source address of this model includes GitHub, Hugging Face, and ModelScope, allowing developers to freely download and experience it. Step3 adopts an MoE architecture, with a total parameter count of 321 billion and an activated parameter count of 38 billion. It

Tencent's X-Omni AI model integrates reinforcement learning to unify image generation and understanding, outperforming rivals in text rendering and visual tasks without classifier guidance.....

The paper by Dr. Liang Wenfeng's team at DeepSeek in collaboration with Peking University has won the ACL 2025 Best Paper Award. This research proposes the Native Sparse Attention (NSA) mechanism, which achieves algorithm-hardware co-optimization through three parallel attention branches: compression, selection, and sliding. It improves the processing speed of long texts by 11 times and outperforms traditional full-attention models. NSA supports a context length of up to 1 million tokens, achieving an 11.6 times decoding speed improvement on 64k text, and a 27B parameter model demonstrates excellent performance on multiple benchmark tests.

Kunlun open-sources Skywork UniPic, a 1.5B-parameter multimodal model integrating image understanding, generation, and editing. It unifies multimodal tasks with end-to-end pretraining, offering strong performance comparable to larger models.....

Google upgrades the Imagen4 image generation model. Imagen4Ultra ranks third globally in authoritative rankings, with performance close to GPT-4o and Seedream3.0. The new version significantly improves image details, realism, and style consistency, and handles complex prompts more accurately. It has a clear price advantage, with the standard version at $40 per thousand images and the Ultra version at $60 per thousand images, far lower than GPT-4o's $167. The generation speed is 9.5 seconds per image, faster than GPT-4o but slightly slower than Seedream3.

Alibaba opensources the video generation model Tongyi Wanxiang Wan2.2, which includes three core models: text-to-video, image-to-video, and unified video generation. Innovations include: 1) the first MoE architecture in the industry, improving computational efficiency by 50%; 2) a cinematic aesthetic control system, enabling precise control over lighting, color, and other cinematic effects; 3) a 5B small-sized unified model that can be deployed on consumer-grade GPUs. This series of models has been open-sourced on platforms such as GitHub, with over 5 million downloads, significantly lowering the barrier to AI video generation.

VLM2Vec-V2, a multimodal embedding framework developed by Salesforce Research and other institutions, breaks through the limitations of traditional models. Based on the Qwen2-VL architecture, this framework innovatively unifies image, video, and visual document retrieval tasks. It adds five new evaluation tasks to expand the MMEB dataset. With key technologies such as dynamic resolution and M-RoPE, it achieved an average score of 58.0 across 78 datasets, leading in performance, especially in video tasks. Although its document retrieval is slightly behind ColPali, it

Alibaba's WebSailor, an open-source AI agent framework, excels in BrowseComp-en/zh and SimpleQA tests. Features knowledge graph-based task generation and reinforcement learning with RFT/DUPO. Introduces data obfuscation for enhanced AI training.....

At the 2025 World Artificial Intelligence Conference (WAIC), Baidu presented multiple innovative achievements and latest developments in the field of artificial intelligence. During the conference, Baidu announced that its Apollo Go, PaddlePaddle Deep Learning Platform, and Baidu Intelligent Computing Cluster have been selected for the China Artificial Intelligence Industry Innovation Exhibition. Among them, Apollo Go not only exhibited as an item but also served as a shuttle vehicle for the conference, demonstrating the mature application of its autonomous driving technology. Currently, Apollo Go has provided over 11 million travel services worldwide, with safety driving mileage exceeding 1

iFlytek's upgraded Spark X1 model enhances translation, reasoning, and text generation, achieving top-tier global performance with 20% better translation quality. It scores over 80 in EN-CN conference translation, reduces AI hallucinations, and boosts efficiency for 100+ firms by 50%. Available via web, app, and API.....

Alibaba Cloud is about to release the upgraded video generation AI model Wan 2.2, achieving multiple breakthroughs based on Wan 2.1: 1) Performance improvements, supporting high-resolution 1080p/4K and long video generation; 2) New video-to-audio function, enhancing multimodal creation; 3) Optimized hardware requirements, capable of running on devices with 6GB VRAM; 4) Training data expanded to 1.5 billion videos + 10 billion images; 5) Continued open-source strategy, using the Apache 2.0 license. This model will solidify Alibaba's leading position in the field of AI video generation.

Google Photos introduces AI features to turn static photos into 6s videos with artistic styles, powered by Veo2 and Imagen models. Features include digital watermarks and are rolling out in the US, with Remix coming soon.....

Kuaishou and Shanghai Jiao Tong University announced the open-source multimodal generation model Orthus at the ICML conference. The model is based on an autoregressive Transformer architecture and demonstrates excellent performance in text-image conversion tasks, with significantly improved computational efficiency compared to existing models. Orthus uses a unique architectural design, separating processing of text and image features through specific modal generation heads, and efficiently fusing them in a unified representation space. Research shows that the model outperforms professional models in image understanding and text-to-image tasks, while also demonstrating strong potential in application scenarios such as image editing, making it a breakthrough in multimodal generation.

Fogsight is an AI animation engine based on LLM that can convert abstract concepts into teaching animations with one click. Users can input keywords to generate short videos with bilingual voiceover and cinematic visual effects, lasting 30-90 seconds, suitable for teaching in subjects such as mathematics and physics. Its core advantage lies in intelligent arrangement throughout the entire process, lowering the production barrier. It supports local deployment and open-source customization, but there is still room for improvement in handling complex scenarios. Compared to competitors like Animaker, Fogsight focuses more on concept-driven generation for educational scenarios. It is already being used in classrooms and online

AMD and Tensorstack jointly launch AI image generation software Amuse 3.1, optimized for devices equipped with Ryzen AI XDNA2NPU. The new version significantly improves image quality and reduces memory usage, allowing SD3.0 medium models to run on 24GB RAM laptops. The software uses a no-subscription model, offering free use for individual users and small businesses. Key features include fast image iteration adjustments, local offline generation, and image resolution doubling through XDNA2NPU technology.

The Qualcomm AI Research Institute has introduced an innovative technology called CSD-VAR, which enhances the generation capabilities of vision autoregressive models through content-style decomposition. This technology employs scale-aware generation and SVD correction to achieve precise separation of content and style, outperforming traditional diffusion models in terms of fidelity and stylistic effects. The accompanying CSD-100 dataset optimizes model training and evaluation. CSD-VAR also introduces a K-V memory mechanism to improve processing efficiency, with broad application prospects in fields such as artistic creation and game development, showcasing Qualcomm's advancements in visual generation technology.

FotographerAI introduces the revolutionary image generation framework ZenCtrl, which enables high-fidelity image generation with multiple perspectives and scenes from a single input image using GenAI technology. This framework breaks through three major pain points of traditional image generation: no need for fine-tuning, maintaining subject consistency, and supporting fine-grained control. Its modular toolkit integrates the entire workflow from preprocessing to postprocessing, built upon and optimized from OmniControl, with particular emphasis on enhancing subject retention capabilities. It supports diverse scenarios such as advertising creativity and virtual try-on, and future developments are also planned.

ComfyUI-Copilot, developed by Alibaba, is an open-source plugin based on large language models that significantly simplifies the AI art creation process through natural language interaction. The tool includes a knowledge base with 7,000 nodes, 62,000 models, and 9,000 workflows, supporting automatic workflow generation, intelligent node recommendation, and parameter optimization. Since its open-source release in February 2025, it has attracted 19,000 users, handled 85,000 queries, and achieved an 85.9% workflow acceptance rate. Its features include multi-language support, error diagnosis