Quanta AI Labs

22 / May 26 – Jun 1, 2025 · 8 RELEASES

AI News Week 22, 2025

Week 22 2025 AI releases.

Vision & Image (1)

  • 4KAgent

    Texas A&M University, Stanford University, and collaborators · Image Restoration

    An agentic image super-resolution system that universally upscales any image to 4K resolution across diverse domains—natural scenes, degraded photos, AIGC, and scientific imaging (medical, remote sensing, microscopy)—without requiring retraining or domain-specific adaptation.

Video & Animation (2)

  • StreamDiT

    Meta / UC Berkeley · Text-to-Video

    Real-time text-to-video generation model achieving 16 FPS on single H100 GPU using flow matching with moving buffer and distillation. Enables interactive streaming video generation at 512p resolution with dynamic prompt-based control.

  • MultiTalk

    Sun Yat-sen University, Meituan, HKUST · Text-to-Video

    Audio-driven multi-person conversational video generation framework that resolves audio-person binding through Label Rotary Position Embedding (L-RoPE) and maintains instruction-following capabilities via partial parameter and multi-task training.

Audio & Speech (1)

  • OmniVCus

    Caiyuan Hao (Research) · Voice Cloning

    A voice conversion system enabling flexible timbre transfer and style control across speakers. Supports various voice conversion tasks with omni-directional capability.

3D & Spatial (2)

  • OmniPart

    University of Hong Kong, VAST, Harbin Institute of Technology, Zhejiang University · 3D Generation

    OmniPart is a framework for part-aware 3D object generation that decouples structure planning from part synthesis using autoregressive bounding box prediction and spatially-conditioned diffusion models. It enables controllable, editable 3D asset creation with explicit part structures from single images, supporting diverse downstream applications like animation, material editing, and geometry processing.

  • SpatialTrackerV2

    Zhejiang University, University of Oxford, Ant Research, Pixelwise AI, Bytedance · 3D Reconstruction

    End-to-end 3D point tracking model that jointly estimates camera motion, scene geometry, and pixel-wise 3D trajectories from monocular video in a fully differentiable architecture. Achieves state-of-the-art performance on 3D tracking while also excelling at 2D tracking and dynamic 3D reconstruction.

Agents & Automation (2)

  • Reachy Mini

    Hugging Face · Robotics

    Reachy Mini is a compact humanoid robot platform with integrated AI capabilities for manipulation and interaction tasks. The release includes model weights and software tools for robotics developers to build autonomous behaviors.

  • LeRobot

    Hugging Face · Robotics

    Open-source framework for robot learning that provides pre-trained models, datasets, and tools for training robots on manipulation and navigation tasks using imitation learning and reinforcement learning approaches.

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.