Quanta AI Labs

17 / Apr 21 – 27, 2025 · 7 RELEASES

AI News Week 17, 2025

Week 17 2025 AI releases.

Vision & Image (2)

  • PixelHacker

    Huazhong University of Science and Technology & VIVO AI Lab · Image Editing

    PixelHacker is a diffusion-based image inpainting model that uses latent categories guidance (LCG) to maintain structural and semantic consistency by encoding foreground/background representations as fixed-size embeddings injected via linear attention during denoising. Trained on a 14M image-mask pair dataset, it achieves state-of-the-art results on Places, CelebA-HQ, and FFHQ benchmarks.

  • T2I-R1

    CaraJ7 · Text-to-Image

    Text-to-image generation model or framework based on GitHub repository structure. Specific technical details unavailable due to page truncation.

Video & Animation (2)

  • HoloTime

    Peking University, Peng Cheng Laboratory, Harbin Institute of Technology · Text-to-Video

    HoloTime generates high-fidelity 360° panoramic videos from text or images using a two-stage diffusion model, then reconstructs them into spatially-temporally consistent 4D scenes via space-time depth estimation and Gaussian Splatting for immersive VR/AR experiences.

  • FlexiAct

    Tsinghua University / ARC Lab, Tencent PCG · Video Editing

    FlexiAct enables flexible action transfer from reference videos to target images across heterogeneous scenarios with varying spatial structures, skeletons, and viewpoints using RefAdapter for spatial adaptation and FAE for frequency-aware action extraction during the denoising process.

3D & Spatial (1)

  • PrimitiveAnything

    Tencent AIPD & Tsinghua University · 3D Generation

    An auto-regressive transformer framework that decomposes complex 3D shapes into primitive assemblies by learning from large-scale human-crafted abstractions, using an ambiguity-free parameterization scheme to represent multiple primitive types uniformly. Demonstrates superior generalization across diverse shape categories while maintaining alignment with human perception.

Agents & Automation (1)

  • ZeroSearch

    Alibaba NLP · Agent Framework

    A search-based framework that enables language models to perform information retrieval and reasoning without explicit training on search tasks. Implements zero-shot search capabilities by leveraging LLM reasoning over retrieved documents.

Developer Tools (1)

  • 3D Guided GenAI RTX

    NVIDIA · AI Infrastructure

    NVIDIA blueprint for 3D-guided generative AI applications optimized for RTX GPUs, enabling developers to build multimodal generative systems that leverage 3D scene understanding and spatial context for enhanced AI outputs.

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