Quanta AI Labs

15 / Apr 7 – 13, 2025 · 10 RELEASES

AI News Week 15, 2025

Week 15 2025 AI releases.

Vision & Image (3)

  • InstantCharacter

    InstantX · Text-to-Image

    A character generation system that creates personalized digital characters, likely using diffusion or neural rendering techniques for rapid character instantiation.

  • Cobra

    Tsinghua University, Chinese University of Hong Kong, Tencent ARC Lab · Image Editing

    Cobra is an efficient line art colorization framework using a Causal Sparse DiT architecture that processes 200+ reference images with low latency while preserving color identity and supporting interactive color hints. Accepted by SIGGRAPH 2025.

  • Step1X-Edit

    Stepfun AI · Image Editing

    Image editing framework that leverages diffusion models for controlled image manipulation and generation tasks. Provides tools for precise editing operations with step-by-step guidance.

Video & Animation (4)

  • Wan 2.1

    Wan-Video · Text-to-Video

    Video generation model supporting extended context and improved temporal consistency. Enables creation of longer, higher-quality video sequences with better scene coherence.

  • Sonic

    Tencent & Zhejiang University · Text-to-Video

    Audio-driven portrait animation system that generates stable long-form videos (1-10 minutes) by disentangling global audio perception into intra-clip and inter-clip components, using context-enhanced learning and time-aware position shift fusion for coherent head movement and lip-sync.

  • Seaweed

    Research Team (Ceyuan Yang et al.) · Text-to-Video

    A 7B-parameter diffusion transformer video generation model trained on 1,000 H100 GPU equivalents that generates videos from text, images, and audio with support for multi-shot long-form narratives, real-time generation at 1280x720, and up to 1-minute sequences.

  • FramePack

    Stanford University, MIT (Lvmin Zhang et al.) · Text-to-Video

    FramePack optimizes video diffusion by dynamically packing frame contexts with variable compression rates and introduces bi-directional anti-drifting sampling to prevent quality degradation in long video generation. Enables 13B video models to run on 6GB laptop GPUs and generate thousands of frames at full fps-30.

3D & Spatial (1)

  • PartField

    NVIDIA Toronto AI Lab · 3D Reconstruction

    PartField is a feedforward neural network that learns part-based 3D feature fields for open-world shape segmentation, achieving 20% higher accuracy and orders of magnitude faster inference than existing methods. The learned features exhibit emergent cross-shape consistency enabling applications like co-segmentation and correspondence without explicit supervision.

Developer Tools (2)

  • ComfyUI-FramePackWrapper

    kijai · AI Infrastructure

    A ComfyUI extension that enables efficient frame batching and packing for video processing workflows, optimizing memory usage and computational performance in generative video tasks.

  • FramePack

    lllyasviel · AI Infrastructure

    FramePack appears to be a framework for efficient frame packing and video processing, enabling optimized handling of sequential visual data through structured packaging mechanisms.

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