Our Blog

ByteDance unveils DreamID-Omni, a unified framework that generates, edits, and animates human-centric videos with synchronized audio, achieving state-of-the-art results that rival commercial models.

Meta AI introduces VecGlypher, a breakthrough multimodal language model that generates high-fidelity vector fonts directly from text descriptions or image examples, revolutionizing digital typography.

We introduce the Sphere Encoder, an efficient generative framework capable of producing images in a single forward pass and competing with many-step d...

New research introduces SeaCache, a training-free acceleration method that makes diffusion models significantly faster by understanding how images evolve from low to high frequencies during generation.

Open-source GUI agents get a major boost with GUI-Libra, a new training framework that combines action-aware supervision and partially verifiable reinforcement learning.

Exploring how large language models store and retrieve knowledge - insights from the NanoKnow: How to Know What Your Language Model Knows research paper.

NoLan addresses object hallucinations in vision-language models by dynamically suppressing language decoder priors, achieving 6-7% accuracy improvements without retraining.

Revolutionary framework aligns wearable sensor data with video to achieve sub-second precision in motion tracking and action recognition.

Retrieval algorithms like BM25 and query likelihood with Dirichlet smoothing remain strong and efficient first-stage rankers, yet improvements have mostly relied on parameter tuning and human intui...

Researchers from University of Wisconsin-Madison introduce TAPE, a framework that dramatically improves AI agent reliability by 21% through better planning and error-resistant execution.

Can AI automatically discover better search algorithms? RankEvolve uses LLMs and evolutionary techniques to create novel retrieval algorithms that outperform traditional methods.

New research on query-focused reranking for long context processing in language models.

Researchers use LLMs and evolutionary search to automatically discover novel retrieval algorithms that outperform traditional methods like BM25.

Researchers introduce SIMSPINE, the first open dataset for 3D spinal motion analysis with over 2 million frames, bridging musculoskeletal simulation and computer vision.

UW researchers unveil ADRA, a groundbreaking technique that uses reinforcement learning to detect what data was used to train AI models, achieving 18.8% improvement over previous methods.

TAPE introduces a novel framework for language model agents that combines tool-guided planning with constrained execution to improve reliability in environments with sequential dependencies.

ArtiAgent introduces an automated approach to detect and fix visual artifacts in AI-generated images using three specialized agents.

ArtiAgent introduces an automated system using three specialized agents to generate and annotate visual artifacts in AI-generated images, creating scalable training data for improved image quality.

Want to opt out of Google's Gemini AI features? Here's a complete guide to disabling Gemini in Gmail, Google Photos, Chrome, and other Google Workspace apps.

We are excited to launch our official blog! Here, we will share insights on technology, business growth, and industrial innovations.