Quanta AI Labs

01 / Dec 30, 2024 – Jan 5, 2025 · 10 RELEASES

AI News Week 1, 2025

Week 1 2025 AI releases.

Video & Animation (6)

  • MatAnyone

    S-Lab (Nanyang Technological University) & SenseTime Research · Video Editing

    A memory-based video matting framework that achieves stable human segmentation with fine-grained boundary details through region-adaptive memory fusion and a novel training strategy leveraging both matting and segmentation data.

  • VideoJAM

    Hila Chefer et al. · Video Editing

    VideoJAM enables zero-shot video editing through joint attention map manipulation, allowing precise control over object appearance and motion without per-video training or manual masks.

  • DiffVSR

    Shanghai Jiao Tong University, Shanghai AI Lab, SIAT · Video Editing

    Diffusion-based video super-resolution method using Progressive Learning Strategy to decompose complex degradation modeling into staged training, and Interweaved Latent Transition for temporal consistency. Achieves superior performance on severely degraded videos where existing methods fail.

  • Pikaformance

    Pika · Text-to-Video

    A web-based AI model that generates hyper-realistic facial expressions and lip-sync animations synchronized to any audio input, enabling images to speak, sing, rap, and produce other vocalizations with near real-time generation speed.

  • MotionCanvas

    Adobe Research, Chinese University of Hong Kong, Monash University · Text-to-Video

    MotionCanvas enables intuitive cinematic shot design in image-to-video generation by allowing joint control of camera movements and object motions in a 3D-scene-aware manner, using spatiotemporal motion-conditioning signals for video diffusion models without requiring 3D training data.

  • OmniHuman-1

    ByteDance Intelligent Creation · Text-to-Video

    End-to-end multimodality-conditioned human video generation framework that synthesizes realistic human videos from a single image and motion signals (audio, video, or combined). Introduces mixed training strategy to leverage data scaling across different conditioning modalities, enabling high-quality output from weak signal inputs.

3D & Spatial (3)

  • Meshcapade Motion Capture Editor

    Meshcapade · 3D Reconstruction

    Markerless motion capture solution using AI to track human movement without physical markers. Enables real-time 3D body pose estimation from video input.

  • Meshcapade Markerless Motion Capture

    Meshcapade GmbH · 3D Reconstruction

    Markerless motion capture platform using single camera input to capture full-body and finger movements with SMPL-based 3D body modeling. Enables real-time pose estimation for gaming, fashion virtual try-on, and robotics with cloud/on-premises deployment options.

  • Hunyuan3D-2

    Tencent · 3D Generation

    A 3D content generation model that creates high-quality 3D assets from text and images. Builds on diffusion-based approaches to enable efficient single-stage 3D generation with improved geometry and texture quality.

Agents & Automation (1)

  • ReasonerAgent

    Maitrix.org · Web Agents

    Open-source web browser agent that performs multi-step research tasks using world model simulation for planning instead of pure autoregressive reasoning, achieving 32.2% success on flight searches (0% for baselines) without requiring fine-tuning or RL training.

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